All-terrain adaptive mobile chassis control method, device, equipment, medium, robot and system

Abstract

The application provides a control method and device of an all-terrain adaptive mobile chassis, electronic equipment, a computer readable storage medium, a mobile robot and a control system. The method obtains a terrain elevation feedback sequence of a plurality of independently driven wheel-leg composite units at a current time, analyzes terrain mutation characteristics, inputs a pre-configured attitude decision network to determine a target vehicle body attitude angle, calculates the swing arm adjustment angle and the telescopic adjustment length of each wheel-leg composite unit, sequentially controls the swing arm action and the telescopic action, and iteratively updates the network mapping parameters according to the chassis balance state characteristics after the telescopic action.

Description

Technical Field

[0001] This application relates to the field of mobile robot chassis control technology, and more specifically, to a control method, device, electronic device, computer-readable storage medium, mobile robot, and control system for an all-terrain adaptive mobile chassis. Background Technology

[0002] With the increasing demand for mobile robots in scenarios such as field inspections, emergency rescues, mining transportation, agricultural operations, and patrols in complex industrial parks, mobile platforms need to maintain good passability and driving stability in complex environments such as steps, gravel, slopes, snow, and soft ground. In these scenarios, changes in ground morphology are often discontinuous and locally abrupt. During movement, mobile platforms are easily affected by changes in local support height, changes in contact point forces, and vehicle posture deviations. Therefore, higher requirements are placed on the adaptability and control response capabilities of the chassis structure.

[0003] A common approach to improving obstacle-crossing capability for mobile platforms is to employ a tracked chassis, a fixed-height frame, and a passive suspension structure. This approach reduces the risk of localized subsidence by increasing the ground contact area, absorbs ground impacts through the passive suspension, and adapts to different road surfaces by using pre-set travel speeds or obstacle-crossing modes. The key control considerations for this type of solution are primarily focused on speed limits, power output distribution, and the selection of passive buffer structures.

[0004] However, the above solutions mainly rely on structural redundancy and preset operating modes to adapt to complex terrain. When the terrain changes abruptly in a local area, it is difficult to incorporate the real-time feedback of multiple grounding areas, environmental change identification, vehicle posture adjustment, actuator action sequence, and post-action balance feedback into the same control link. This results in a lag in chassis posture adjustment, and changes in local support can easily cause vehicle tilting, jamming, or decreased stability. It is difficult to meet the requirements of all-terrain mobile platforms for adaptive passability and closed-loop control capability under abrupt terrain changes. Summary of the Invention

[0005] This application provides a control method, apparatus, electronic device, computer-readable storage medium, mobile robot, and control system for an all-terrain adaptive mobile chassis, to at least alleviate the aforementioned technical problems.

[0006] A control method for an all-terrain adaptive mobile chassis includes: Obtain the terrain elevation feedback sequence of multiple independently driven wheel-leg composite units on the all-terrain adaptive mobile chassis at the current moment; Spatial distribution analysis is performed on the terrain elevation feedback sequence to extract the terrain abrupt change features of the current environment; The terrain elevation feedback sequence and the terrain abrupt change features are input into a pre-configured attitude decision network to determine the target vehicle attitude angle of the all-terrain adaptive mobile chassis. Based on the target vehicle body attitude angle, calculate the swing arm adjustment angle and telescopic adjustment length of each of the multiple independently driven wheel-leg composite units; The corresponding wheel-leg composite unit is controlled to perform the swing arm action according to the swing arm adjustment angle, and after the swing arm action is completed, the corresponding wheel-leg composite unit is controlled to perform the telescopic action according to the telescopic adjustment length. The network mapping parameters in the attitude decision network are iteratively updated based on the chassis balance state characteristics after performing the above-mentioned telescopic action.

[0007] Optionally, the step of performing spatial distribution analysis on the terrain elevation feedback sequence to extract terrain abrupt change features of the current environment includes: The terrain elevation feedback sequence is mapped to a preset chassis global coordinate system to generate a continuous terrain point cloud outline that represents the terrain undulation state at the current moment. The spatial distribution is analyzed in the continuous terrain point cloud outline to extract the elevation change gradient between adjacent coordinate points; The coordinate points where the elevation change gradient exceeds a preset gradient threshold are marked as abrupt change candidate points; The multiple mutation candidate points are subjected to regional feature aggregation processing to generate the terrain mutation feature containing the spatial distribution density and contour geometry of the mutation candidate points.

[0008] Optionally, the step of inputting the terrain elevation feedback sequence and the terrain abrupt change features into a pre-configured attitude decision network to determine the target vehicle attitude angle of the all-terrain adaptive mobile chassis includes: The terrain elevation feedback sequence is temporally featured by the state-aware coding layer in the attitude decision network to obtain the first hidden state representation. The spatial attention layer in the posture decision network is used to assign environmental obstacle weights to the terrain abruptness features to obtain the second hidden state representation. The first hidden state representation and the second hidden state representation are concatenated, and then the concatenated features are input into the policy output layer of the attitude decision network to generate the target vehicle attitude angle of the all-terrain adaptive mobile chassis to maintain a baseline level state under the corresponding terrain.

[0009] Optionally, the step of calculating the swing arm adjustment angle and telescopic adjustment length corresponding to each of the plurality of independently driven wheel-leg composite units based on the target vehicle body attitude angle includes: Obtain the initial physical configuration parameters of the multiple independently driven wheel-leg composite units at the current moment; Construct the inverse kinematics transformation matrix of the all-terrain adaptive mobile chassis based on the target vehicle body attitude angle; Substitute the initial physical configuration parameters into the inverse kinematics transformation matrix to obtain the expected relative coordinate positions of each wheel-leg composite unit relative to the all-terrain adaptive mobile chassis. Based on the desired relative coordinate position, the expected rotation value of the rotary joint and the expected stroke value of the linear actuator corresponding to each of the wheel-leg composite units to reach the coordinate position are calculated respectively. Then, the expected rotation value of the rotary joint is used as the adjustment angle of the swing arm, and the expected stroke value of the linear actuator is used as the extension adjustment length.

[0010] Optionally, the step of calculating the expected rotation value of the rotary joint and the expected stroke value of the linear actuator corresponding to the expected relative coordinate position for each of the wheel-leg composite units includes: For each of the wheel-leg composite units, the corresponding grounding contact position is extracted based on the desired relative coordinate position to establish a first kinematic geometric constraint condition that reflects the relationship between the swing arm adjustment angle and the corresponding grounding contact position. Establish a second kinematic geometric constraint condition that reflects the relationship between the telescopic adjustment length and the position of the grounding contact; Under the premise of satisfying the first kinematic geometric constraints and the second kinematic geometric constraints, the optimization objective is to minimize the overall energy consumed by the wheel-leg composite unit in performing the swing arm movement and the extension movement. Then, within the feasible domain of the joint, the expected value of the rotation of the rotary joint and the expected value of the stroke of the linear actuator are calculated.

[0011] Optionally, the step of controlling the corresponding wheel-leg composite unit to perform the swing arm movement according to the swing arm adjustment angle, and controlling the corresponding wheel-leg composite unit to perform the telescopic movement according to the telescopic adjustment length after the swing arm movement is completed, includes: A first drive command stream is generated based on the swing arm adjustment angle to instruct the swing arm movement, and then the first drive command stream is sent to the rotary joint actuator corresponding to the wheel-leg composite unit to control it to perform the swing arm movement. The angle encoder within the wheel-leg composite unit monitors the swing arm's execution angle in real time. When it is determined that the swing arm execution angle has reached the swing arm adjustment angle and the swing arm action has been completed, a second drive command stream is generated based on the telescopic adjustment length to indicate the telescopic action. The second drive command stream is then sent to the linear drive actuator corresponding to the wheel-leg composite unit to control it to execute the telescopic action.

[0012] Optionally, the iterative update of the network mapping parameters in the attitude decision network based on the chassis balance state characteristics after performing the above-mentioned telescopic action includes: The inertial measurement unit installed on the all-terrain adaptive mobile chassis collects the current pitch angle and current roll angle of the chassis after performing the above-mentioned telescopic action. Based on the first deviation value of the current pitch angle of the chassis from the horizontal plane and the second deviation value of the current roll angle of the chassis from the horizontal plane, an environmental interaction reward value is constructed, and then the environmental interaction reward value is used as the chassis balance state feature. For the attitude decision network, a time difference error is calculated based on the environmental interaction reward value, which is a feature of the chassis balance state, and then gradient information is calculated based on the time difference error. The network mapping parameters in the attitude decision network are iteratively updated according to the descent direction corresponding to the gradient information in order to maximize the cumulative environmental interaction reward value generated by subsequent actions.

[0013] Optionally, the step of constructing the environmental interaction reward value based on the first deviation value of the current pitch angle of the chassis from the horizontal plane and the second deviation value of the current roll angle of the chassis from the horizontal plane includes: The first deviation value, which represents the current pitch angle of the chassis deviating from the horizontal plane, and the second deviation value, which represents the current roll angle of the chassis deviating from the horizontal plane, are respectively reverse-mapped to obtain a first basic reward item and a second basic reward item that are inversely proportional to the degree of deviation. The peak value of the transient torque change fed back by each of the wheel-leg composite units during the above-mentioned extension and retraction action is obtained, and then a penalty term for suppressing the intensity of the action is constructed based on the peak value of the transient torque change. The first basic reward item and the second basic reward item are fused to obtain the basic reward index; By introducing the penalty term to reduce the weight of the basic reward indicator, the environmental interaction reward value is constructed based on the penalty term after the weight reduction.

[0014] Optionally, the terrain elevation feedback sequence includes local detection distance data acquired by the multiple wheel-leg composite units respectively through their built-in non-contact ranging probes; acquiring the terrain elevation feedback sequence of the multiple independently driven wheel-leg composite units on the all-terrain adaptive mobile chassis at the current moment includes: At the current moment, the non-contact ranging probes built into the multiple independently driven wheel-leg composite units on the all-terrain adaptive mobile chassis are synchronously triggered to perform local ground ranging. Acquire the local detection distance data returned by each of the non-contact ranging probes; The local detection distance data are sorted and combined according to the structural installation topology of the wheel-leg composite unit on the chassis to form a terrain elevation feedback sequence covering the grounding area around the all-terrain adaptive mobile chassis.

[0015] A control device for an all-terrain adaptive mobile chassis includes: The feedback acquisition module is used to acquire the terrain elevation feedback sequence of multiple independently driven wheel-leg composite units on the all-terrain adaptive mobile chassis at the current moment. The feature extraction module is used to perform spatial distribution analysis on the terrain elevation feedback sequence in order to extract the terrain abrupt change features of the current environment; The attitude determination module is used to input the terrain elevation feedback sequence and the terrain abrupt change features into a pre-configured attitude decision network to determine the target vehicle attitude angle of the all-terrain adaptive mobile chassis. The adjustment parameter calculation module is used to calculate the swing arm adjustment angle and telescopic adjustment length of each of the multiple independently driven wheel-leg composite units according to the target vehicle body attitude angle. The motion control module is used to control the corresponding wheel-leg composite unit to perform the swing arm motion according to the swing arm adjustment angle, and after the swing arm motion is completed, control the corresponding wheel-leg composite unit to perform the telescopic motion according to the telescopic adjustment length. The parameter update module is used to iteratively update the network mapping parameters in the attitude decision network based on the chassis balance state characteristics after performing the above-mentioned telescopic action.

[0016] An electronic device, comprising: At least one processor; and, A memory that is communicatively connected to the at least one processor; The memory stores a computer program that can be executed by the at least one processor. When the computer program is executed by the at least one processor, it causes the at least one processor to perform the control method for the all-terrain adaptive mobile chassis as described in this application.

[0017] A computer-readable storage medium storing a computer program; When the computer program is executed by the processor, it implements the control method for the all-terrain adaptive mobile chassis as described in this application.

[0018] A mobile robot, comprising: The robot itself; An all-terrain adaptive mobile chassis is mounted on the robot body, and the all-terrain adaptive mobile chassis includes multiple independently driven wheel-leg composite units; The terrain elevation feedback acquisition component is used to acquire the terrain elevation feedback sequence of the multiple independently driven wheel-leg composite units at the current moment; The chassis controller is communicatively connected to the terrain elevation feedback acquisition component and the multiple independently driven wheel-leg composite units; The chassis controller is used to perform spatial distribution analysis on the terrain elevation feedback sequence in order to extract the terrain abrupt change features of the current environment; The chassis controller is also used to input the terrain elevation feedback sequence and the terrain abrupt change features into a pre-configured attitude decision network to determine the target vehicle attitude angle of the all-terrain adaptive mobile chassis. The chassis controller is also used to calculate the swing arm adjustment angle and telescopic adjustment length of each of the multiple independently driven wheel-leg composite units according to the target vehicle body attitude angle. The chassis controller is also used to control the corresponding wheel-leg composite unit to perform the swing arm action according to the swing arm adjustment angle, and after the swing arm action is completed, control the corresponding wheel-leg composite unit to perform the telescopic action according to the telescopic adjustment length. The chassis controller is also used to iteratively update the network mapping parameters in the attitude decision network based on the chassis balance state characteristics after performing the above-mentioned telescopic action.

[0019] A control system for an all-terrain adaptive mobile chassis includes: An all-terrain adaptive mobile chassis, comprising multiple independently driven wheel-leg composite units; A terrain elevation feedback acquisition device is used to acquire the terrain elevation feedback sequence of the multiple independently driven wheel-leg composite units at the current moment; The attitude control device is communicatively connected to the terrain elevation feedback acquisition device and the multiple independently driven wheel-leg composite units; The attitude control device is used to perform spatial distribution analysis on the terrain elevation feedback sequence in order to extract the terrain abrupt change features of the current environment. The attitude control device is also used to input the terrain elevation feedback sequence and the terrain abrupt change features into a pre-configured attitude decision network to determine the target vehicle attitude angle of the all-terrain adaptive mobile chassis. The attitude control device is also used to calculate the swing arm adjustment angle and telescopic adjustment length of each of the multiple independently driven wheel-leg composite units according to the target vehicle body attitude angle. The attitude control device is also used to control the corresponding wheel-leg composite unit to perform the swing arm action according to the swing arm adjustment angle, and after the swing arm action is completed, control the corresponding wheel-leg composite unit to perform the telescopic action according to the telescopic adjustment length. The attitude control device is also used to iteratively update the network mapping parameters in the attitude decision network based on the chassis balance state characteristics after performing the above-mentioned telescopic action.

[0020] The technical advantages of the technical solution provided in this application are: This application presents a control method for an all-terrain adaptive mobile chassis. Addressing the shortcomings of existing mobile platforms in integrating real-time ground contact feedback, environmental change identification, vehicle attitude adjustment, actuator sequence, and post-action balance feedback into a single control link under abrupt terrain changes, this method first acquires the terrain elevation feedback sequence of multiple independently driven wheel-leg composite units on the all-terrain adaptive mobile chassis at the current moment. Then, it performs spatial distribution analysis on the terrain elevation feedback sequence to extract the terrain change characteristics of the current environment. Compared to traditional methods that rely solely on fixed structural redundancy or preset operating modes, this method allows the chassis control process to use the feedback from multiple wheel-leg composite units at the current moment as input, and further incorporates environmental change information into the judgment criteria before attitude adjustment. This helps reduce the attitude adjustment lag problem caused by the failure to promptly incorporate local terrain changes into the control link.

[0021] Based on this, this application inputs the terrain elevation feedback sequence and terrain abrupt change characteristics into a pre-configured attitude decision network to determine the target vehicle attitude angle of the all-terrain adaptive mobile chassis. Then, based on the target vehicle attitude angle, it calculates the corresponding swing arm adjustment angle and telescopic adjustment length for each of the multiple independently driven wheel-leg composite units. Compared to traditional solutions that passively adapt to ground changes through a unified operating mode or passive suspension structure, this application establishes a continuous technical interaction relationship between terrain feedback, abrupt change characteristics, the target vehicle attitude, and the adjustment parameters of each wheel-leg composite unit. This allows vehicle attitude adjustment to no longer rely solely on structural buffering, but rather to form a computational link from environmental conditions to target attitude, and then to the adjustment amounts of each wheel-leg composite unit, thereby improving the chassis's attitude adaptability in complex terrain.

[0022] Furthermore, this application controls the corresponding wheel-leg composite unit to perform the swing arm movement according to the swing arm adjustment angle, and after completing the swing arm movement, controls the corresponding wheel-leg composite unit to perform the telescopic movement according to the telescopic adjustment length. Compared with the traditional processing method of triggering multiple execution actions simultaneously or performing only a single height adjustment, this sequential control relationship distinguishes the swing arm movement and the telescopic movement into two consecutive execution stages, so that the wheel-leg composite unit first completes the swing arm position adjustment, and then performs telescopic compensation according to the telescopic adjustment length. This helps to reduce the attitude fluctuations caused by mutual interference between execution actions, and makes the chassis's action execution process under abrupt terrain have a clearer sequential logic and a more stable support adjustment process.

[0023] Furthermore, this application iteratively updates the network mapping parameters in the attitude decision network based on the chassis balance state characteristics after performing the aforementioned telescopic action. This ensures that the control link does not remain at a single attitude decision and action execution, but rather feeds back the chassis balance state after the action to the parameter update process of the attitude decision network. Compared to the traditional preset mode, which struggles to adjust the control strategy based on actual results, this closed-loop update method allows subsequent attitude decisions to gradually absorb the balance state feedback from the previous action. This improves the control adaptability of the all-terrain adaptive mobile chassis in continuous complex terrain and helps alleviate problems such as vehicle tilt, jamming, or decreased stability. Attached Figure Description

[0024] Figure 1 This application provides a control scenario for an all-terrain adaptive mobile chassis. Figure 2 This application provides a control method for an all-terrain adaptive mobile chassis. Figure 3 This application provides a control device for an all-terrain adaptive mobile chassis. Figure 4 An electronic device is described in an embodiment of this application; Figure 5 This application provides an embodiment of a computer-readable storage medium. Figure 6 This application provides an embodiment of a mobile robot. Figure 7 This application provides a control system for an all-terrain adaptive mobile chassis. Detailed Implementation

[0025] like Figure 1 The image shows a control scenario for an all-terrain adaptive mobile chassis according to an embodiment of this application; as shown... Figure 2 The image shows a control method for an all-terrain adaptive mobile chassis according to an embodiment of this application, comprising: Obtain the terrain elevation feedback sequence of multiple independently driven wheel-leg composite units on the all-terrain adaptive mobile chassis at the current moment; Spatial distribution analysis is performed on the terrain elevation feedback sequence to extract the terrain abrupt change features of the current environment; The terrain elevation feedback sequence and the terrain abrupt change features are input into a pre-configured attitude decision network to determine the target vehicle attitude angle of the all-terrain adaptive mobile chassis. Based on the target vehicle body attitude angle, calculate the swing arm adjustment angle and telescopic adjustment length of each of the multiple independently driven wheel-leg composite units; The corresponding wheel-leg composite unit is controlled to perform the swing arm action according to the swing arm adjustment angle, and after the swing arm action is completed, the corresponding wheel-leg composite unit is controlled to perform the telescopic action according to the telescopic adjustment length. The network mapping parameters in the attitude decision network are iteratively updated based on the chassis balance state characteristics after performing the above-mentioned telescopic action.

[0026] 2. The control method for an all-terrain adaptive mobile chassis according to claim 1, characterized in that, the step of performing spatial distribution analysis on the terrain elevation feedback sequence to extract the terrain abrupt change features of the current environment includes: The terrain elevation feedback sequence is mapped to a preset chassis global coordinate system to generate a continuous terrain point cloud outline that represents the terrain undulation state at the current moment. The spatial distribution is analyzed in the continuous terrain point cloud outline to extract the elevation change gradient between adjacent coordinate points; The coordinate points where the elevation change gradient exceeds a preset gradient threshold are marked as abrupt change candidate points; The multiple mutation candidate points are subjected to regional feature aggregation processing to generate the terrain mutation feature containing the spatial distribution density and contour geometry of the mutation candidate points.

[0027] Preferably, the specific implementation process of the step "mapping the terrain elevation feedback sequence to a preset chassis global coordinate system to generate a continuous terrain point cloud profile representing the terrain undulation state at the current moment" is as follows: Before the terrain elevation feedback sequence at the current moment enters the spatial distribution analysis, the local detection distance data corresponding to each of the multiple independently driven wheel-leg composite units is read, and the structural installation topology relationship of the multiple independently driven wheel-leg composite units on the all-terrain adaptive mobile chassis is read, so that each local detection distance data can correspond to a wheel-leg composite unit installation node, a detection orientation, and a grounding area; the wheel-leg composite unit installation node is derived from the structural installation topology relationship, and the wheel-leg composite unit installation node is used to represent the installation position of the corresponding wheel-leg composite unit on the all-terrain adaptive mobile chassis. The wheel-leg composite unit installation node subsequently participates in the mapping processing of local detection distance data to the preset chassis global coordinate system. The technical essence of the terrain elevation feedback sequence is to form terrain height status data by orderly arranging the local detection distance data collected by multiple independently driven wheel-leg composite units at the same current moment according to the structural installation topology. The terrain elevation feedback sequence is not an isolated distance value of a single sensor point, but a time-level data sequence reflecting the height distribution differences of the ground contact area around the all-terrain adaptive mobile chassis. The time-level data sequence is used to keep multiple local detection distance data at the same current moment consistent in terms of time. The time-level data sequence is then used for spatial distribution analysis, so that the all-terrain adaptive mobile chassis can obtain the terrain undulation state at the current moment before generating the target vehicle attitude angle. Thus, the protection subject of the control method has a technical connection from the terrain elevation feedback sequence to the target vehicle attitude angle. The preset global coordinate system of the chassis is configured when the all-terrain adaptive mobile chassis completes structural assembly calibration or starts initial calibration. During configuration, the projection position of the chassis geometric center onto the horizontal support surface is used as the coordinate reference point. The longitudinal direction of travel of the all-terrain adaptive mobile chassis is used as the longitudinal coordinate direction, the left-right arrangement direction of the multiple independently driven wheel-leg composite units is used as the lateral coordinate direction, and the direction perpendicular to the chassis support plane is used as the elevation coordinate direction. The installation offset of the multiple independently driven wheel-leg composite units, the installation height of the non-contact ranging probe, the detection direction of the non-contact ranging probe, and the structural installation topology are all incorporated into the same coordinate transformation caliber. This coordinate transformation caliber is used to define the transformation reference used when converting local detection distance data from the wheel-leg composite unit installation node to terrain coordinate points under the preset global coordinate system of the chassis. This coordinate transformation caliber is subsequently used in the mapping processing of each local detection distance data point.During mapping, the installation node of the wheel-leg composite unit to which each local detection distance data belongs is first determined based on the structural installation topology. Then, the corresponding horizontal coordinate position is determined based on the installation offset of the wheel-leg composite unit installation node in the preset chassis global coordinate system. Subsequently, the elevation coordinate position at this horizontal coordinate position is calculated based on the installation height of the non-contact ranging probe and the local detection distance data, thus converting each local detection distance data into a terrain coordinate point in the preset chassis global coordinate system. After completing the coordinate transformation of all local detection distance data, multiple terrain coordinate points are spatially connected according to the structural installation topology. Interpolation based on adjacent elevation trends is used to fill in the gaps between adjacent terrain coordinate points that lack direct sampling, thereby generating a continuous terrain point cloud outline. The adjacent terrain coordinate points are derived from multiple spatially connected terrain coordinate points, and these adjacent terrain coordinate points are subsequently used as adjacent coordinate points to participate in the extraction of elevation change gradients. The technical essence of the continuous terrain point cloud contour is that it is a spatialized terrain data that represents the current terrain undulation state under a preset chassis global coordinate system. The continuous terrain point cloud contour transforms the terrain elevation feedback sequence, which was originally arranged according to multiple independently driven wheel leg composite units, into contour data with a correspondence between horizontal coordinate position and elevation coordinate position. Subsequent spatial distribution analysis can directly identify local height change areas in the continuous terrain point cloud contour, thereby providing a unified coordinate basis for the extraction of terrain change features.

[0028] Preferably, in the specific technical implementation of the step "performing spatial distribution analysis in the continuous terrain point cloud contour to extract the elevation change gradient between adjacent coordinate points", the reason for performing spatial distribution analysis in the continuous terrain point cloud contour is that the continuous terrain point cloud contour has already converted the local detection distance data in the terrain elevation feedback sequence into terrain coordinate points with spatial positional relationships in the preset chassis global coordinate system. Spatial distribution analysis can compare the height changes between different grounding areas under the same coordinate transformation caliber, without mistakenly taking the installation height difference, detection direction difference, or arrangement position difference of different wheel leg composite unit installation nodes as terrain changes. When performing spatial distribution analysis, each terrain coordinate point in the continuous terrain point cloud contour is first read, and an adjacency determination caliber is established based on the structural installation topology relationship and the horizontal coordinate position of the terrain coordinate point in the preset chassis global coordinate system. The adjacency determination caliber is used to limit which terrain coordinate points can be used as adjacent coordinate points to enter the elevation change gradient extraction process. The adjacency determination caliber subsequently calls the structural installation topology relationship and the horizontal coordinate position to avoid determining adjacent coordinate points only according to the data storage order. The technical essence of adjacent coordinate points is that they are two terrain coordinate points that are horizontally continuous in the preset global coordinate system of the chassis and can reflect the same travel area or adjacent grounding area in the structural installation topology, rather than two values ​​that are simply adjacent in the data storage order. When determining adjacent coordinate points, candidate adjacent coordinate points are first screened based on the front-to-back and left-to-right arrangement of multiple independently driven wheel-leg composite units on the all-terrain adaptive mobile chassis. Then, based on whether the horizontal interval between the candidate adjacent coordinate points falls within the adjacency range matching the grounding area of ​​the wheel-leg composite unit, the adjacent coordinate points participating in the elevation change gradient extraction are determined. This ensures that the adjacent coordinate points have both spatial continuity and reflect the ground undulation relationship in the actual path traversed by the all-terrain adaptive mobile chassis. The candidate adjacent coordinate points originate from terrain coordinate points in the continuous terrain point cloud outline. After the candidate adjacent coordinate points are screened for adjacency range, they are formed into adjacent coordinate points, which are subsequently used to calculate the elevation difference and horizontal interval. Subsequently, for each pair of adjacent coordinate points, the elevation coordinates of the two terrain coordinate points are extracted, and the elevation difference between the two elevation coordinates is calculated. Simultaneously, the horizontal interval between the two terrain coordinate points in the preset global coordinate system of the chassis is extracted. Based on this, the elevation difference is scaled according to the horizontal interval to form an elevation change gradient. The technical essence of the elevation change gradient is a numerical representation of the degree of terrain height change per unit horizontal distance. This elevation change gradient differs from directly using the elevation difference because the horizontal interval between different adjacent coordinate points may be different, and directly comparing the elevation difference would lead to incomparable terrain abrupt changes between areas with larger and smaller intervals. After scale normalization through the horizontal interval, the elevation change gradient can reflect the intensity of terrain undulation changes at the same spatial scale.The extracted elevation change gradient continues to serve as the basis for judging the candidate points of abrupt changes, so that the terrain abrupt change recognition is no longer limited to the judgment of single-point elevation change, but forms a continuous technical link with the spatial distribution relationship in the continuous terrain point cloud outline, the relative position relationship of adjacent coordinate points, and the actual grounding area relationship of the all-terrain adaptive mobile chassis.

[0029] Preferably, the specific implementation process of the step "marking the coordinate points where the elevation change gradient exceeds the preset gradient threshold as a mutation candidate point" is as follows: After the elevation change gradient between adjacent coordinate points has been formed, the preset gradient threshold corresponding to the all-terrain adaptive mobile chassis is read first, and the preset gradient threshold is established in correspondence with the preset chassis global coordinate system, the grounding capability of multiple independently driven wheel-leg composite units, and the attitude adjustment range of the all-terrain adaptive mobile chassis. The preset gradient threshold is not an arbitrarily given fixed empirical value. Its pre-configuration method is as follows: During factory calibration or application scenario initialization of the all-terrain adaptive mobile chassis, the maximum allowable swing arm adjustment angle range, the maximum allowable telescopic adjustment length range, the allowable pitch deviation range of the all-terrain adaptive mobile chassis in a reference horizontal state, the allowable roll deviation range of the all-terrain adaptive mobile chassis in a reference horizontal state, and the distance resolution capability of the non-contact ranging probe are read. These parameters are then converted into elevation change boundaries that can be actively compensated by the multiple independently driven wheel-leg composite units in a preset chassis global coordinate system. A decision margin is then set based on these elevation change boundaries to form the preset gradient threshold. The elevation change boundaries characterize the range of terrain height changes that the multiple independently driven wheel-leg composite units can compensate for through swing arm and telescopic movements. The elevation change boundaries and the decision margin jointly participate in the formation of the preset gradient threshold, ensuring that the preset gradient threshold corresponds to the actual attitude adjustment capability of the all-terrain adaptive mobile chassis. The technical essence of the preset gradient threshold is to distinguish between normal terrain undulations and local height abrupt change regions that need to participate in attitude adjustment decisions. The preset gradient threshold is subsequently used to transform the elevation change gradient into abrupt change candidate point that can enter the terrain abrupt change feature construction process. Therefore, the preset gradient threshold and the determination of the target vehicle attitude angle, the calculation of the swing arm adjustment angle, and the calculation of the telescopic adjustment length in the protection theme are all interconnected. During marking, for each group of adjacent coordinate points, the elevation change gradient corresponding to the group of adjacent coordinate points is compared with the preset gradient threshold in the same dimension. When the elevation change gradient does not exceed the preset gradient threshold, the corresponding coordinate point is kept as a normal undulation coordinate point, and the normal undulation coordinate point continues to be retained in the continuous terrain point cloud outline to maintain the continuity of the overall terrain outline. When the elevation change gradient exceeds the preset gradient threshold, the two adjacent coordinate points that generated the elevation change gradient are read, and based on the elevation change direction, the travel direction of multiple independently driven wheel leg composite units, and the grounding area position of the two adjacent coordinate points, the coordinate point that better represents the entrance position of the local height abrupt change region is determined, and the coordinate point is marked as a abrupt change candidate point.The technical essence of the mutation candidate points is that they are spatial location markers identified in the continuous terrain point cloud contour that need to enter the terrain mutation feature construction process. These candidate points retain both the horizontal and vertical coordinates of the coordinate points and carry the judgment result that the elevation change gradient exceeds a preset gradient threshold. This is because the target vehicle attitude angle does not need to respond equally to all conventional undulating coordinate points, but should be primarily affected by local height abrupt change regions. By marking mutation candidate points, local height abrupt change regions in the continuous terrain point cloud contour are transformed into candidate data that can participate in regional feature aggregation processing, enabling subsequent terrain mutation features to be formed around spatial locations that have a significant impact on chassis attitude.

[0030] Preferably, the specific implementation process of the step "performing regional feature aggregation processing on multiple candidate mutation points to generate the terrain mutation features including the spatial distribution density and contour geometry of the candidate mutation points" is as follows: After multiple candidate mutation points in the continuous terrain point cloud contour have been marked, the horizontal coordinate position, elevation coordinate position, elevation change gradient, and correspondence with the grounding area of ​​multiple independently driven wheel-leg composite units of each candidate mutation point in the preset chassis global coordinate system are read first, and regional feature aggregation processing is performed based on the horizontal coordinate position, elevation coordinate position, elevation change gradient, and correspondence with the grounding area of ​​multiple independently driven wheel-leg composite units of each candidate mutation point. The essence of the regional feature aggregation processing technique is to organize multiple dispersed abrupt change candidate points into regionalized data that reflects the morphology of local height abrupt change areas based on spatial proximity and elevation change direction, rather than simply counting the number of abrupt change candidate points. The aim is to enable terrain abrupt change features to describe the distribution of local height abrupt change areas in the ground-based areas in front of, beside, or around the all-terrain adaptive mobile chassis, allowing the subsequent attitude decision network to simultaneously receive information on the strength and distribution pattern of terrain abrupt changes. The strength information of terrain abrupt changes originates from the elevation change gradient, while the distribution pattern information originates from the spatial distribution density and contour geometry of the abrupt change candidate points. Both the strength and distribution pattern information contribute to the formation of terrain abrupt change features. In the specific process of regional feature aggregation, the horizontal coordinates of the candidate abrupt change points are used as the basis for merging candidate abrupt change points whose horizontal distances fall within the adjacent range and whose elevation change directions are consistent, forming candidate abrupt change point regions. These regions originate from the merging of multiple candidate abrupt change points and are subsequently used to calculate the spatial distribution density of candidate abrupt change points and construct their outline geometry. Then, for each candidate abrupt change point region, the density correspondence between the number of candidate abrupt change points within that region and the corresponding ground cover area is calculated to form the spatial distribution density of candidate abrupt change points. The technical essence of this spatial distribution density is data representing the degree of concentration of terrain abrupt changes within a local height change region. This density reflects the influence range of densely gravelly areas, step edge areas, or slope polygonal areas on the attitude adjustment of the all-terrain adaptive mobile chassis. Subsequently, the outer edge coordinate points are extracted along the outer edge of the mutation candidate point region, and the contour geometry is constructed based on the outer edge coordinate point sequence. The outer edge coordinate point sequence is derived from the arrangement relationship of the outer edge coordinate points in the mutation candidate point region, and the outer edge coordinate point sequence is subsequently used to determine the boundary extension direction of the contour geometry.The technical essence of the contour geometry is geometric description data representing the shape orientation, boundary extension direction, and width variation of the abrupt change candidate point region. The contour geometry can distinguish different terrain morphologies such as the edges of lateral steps, the edges of longitudinal trenches, and local protruding obstacles. After extracting the spatial distribution density and contour geometry of the abrupt change candidate points, the spatial distribution density and contour geometry are associated and encapsulated according to the same abrupt change candidate point region, and the positional relationship of this region in the preset chassis global coordinate system is preserved to generate terrain abrupt change features. The technical essence of these terrain abrupt change features is composite spatial description data extracted from the continuous terrain point cloud contour, capable of characterizing the intensity, range, and morphology of terrain abrupt changes at the current moment. This terrain abrupt change feature is subsequently fed into the attitude decision network along with the terrain elevation feedback sequence, enabling the determination of the target vehicle attitude angle to simultaneously consider the overall terrain undulation state and the morphology of local height abrupt change regions. This, in turn, provides a data source that better meets the requirements for traversing abrupt terrain changes in the calculation of the swing arm adjustment angle and telescopic adjustment length.

[0031] Optionally, the step of inputting the terrain elevation feedback sequence and the terrain abrupt change features into a pre-configured attitude decision network to determine the target vehicle attitude angle of the all-terrain adaptive mobile chassis includes: The terrain elevation feedback sequence is temporally featured by the state-aware coding layer in the attitude decision network to obtain the first hidden state representation. The spatial attention layer in the posture decision network is used to assign environmental obstacle weights to the terrain abruptness features to obtain the second hidden state representation. The first hidden state representation and the second hidden state representation are concatenated, and then the concatenated features are input into the policy output layer of the attitude decision network to generate the target vehicle attitude angle of the all-terrain adaptive mobile chassis to maintain a baseline level state under the corresponding terrain.

[0032] Preferably, the specific implementation process of the step "encoding the terrain elevation feedback sequence with temporal features through the state-aware coding layer in the attitude decision network to obtain the first hidden state representation" is as follows: Before the terrain elevation feedback sequence and terrain change features enter the pre-configured attitude decision network, the network configuration record corresponding to the pre-configured attitude decision network is read. The network configuration record is configured before the all-terrain adaptive mobile chassis enters the operation according to the structural installation topology of multiple independently driven wheel-leg composite units, the data arrangement caliber of the terrain elevation feedback sequence, the order of the input fields of the pre-configured attitude decision network, and the order of the output fields of the target vehicle body attitude angle. The network configuration record is used to enable the state-aware coding layer, spatial attention layer, and strategy output layer in the pre-configured attitude decision network to read the terrain elevation feedback sequence and terrain change features according to the same data arrangement caliber, so as to avoid the same local detection distance data being interpreted as the height feedback of different grounding areas between different structural layers. The technical essence of the state-aware coding layer is that it is a data processing structure layer in a pre-configured attitude decision network used to convert the terrain elevation feedback sequence into a data processing structure that can express the continuous changes in terrain undulation at the current moment. It does not simply read the individual values ​​in the terrain elevation feedback sequence, but rather encodes the terrain elevation feedback sequence according to its current arrangement, the structural installation topology of multiple independently driven wheel-leg composite units, and the changing relationships between terrain elevation feedback sequences at adjacent moments. Specifically, during the temporal feature encoding, the input fields of the terrain elevation feedback sequence are first validated based on the network configuration records to form a field-aligned terrain elevation feedback sequence. The field-aligned terrain elevation feedback sequence continues to retain the local detection distance data corresponding to each wheel-leg composite unit, and each local detection distance data is mapped to a wheel-leg composite unit installation node within the state-aware coding layer. The wheel-leg composite unit installation node originates from the structural installation topology of multiple independently driven wheel-leg composite units. The wheel-leg composite unit installation node is used to ensure that the field-aligned terrain elevation feedback sequence maintains the grounding area correspondence during subsequent temporal unfolding. Subsequently, the state-aware coding layer performs temporal unfolding on the field-aligned terrain elevation feedback sequence, extracts the differences in the same field between the current terrain elevation feedback sequence and the terrain elevation feedback sequence of the previous time, to form temporal fluctuation difference data; the temporal fluctuation difference data is used to express the height change trend of the grounding area of ​​the same wheel leg composite unit in continuous sampling time, and the temporal fluctuation difference data is further associated with the field-aligned terrain elevation feedback sequence to form the first hidden state representation.The technical essence of the first hidden state representation is the vehicle attitude forward perception data formed by encoding the terrain elevation feedback sequence with temporal features. The vehicle attitude forward perception data includes the current height state of the grounding area of ​​each wheel leg composite unit, the height change trend at adjacent time moments, and the structural arrangement relationship between different wheel leg composite units. The first hidden state representation is subsequently used in feature stitching, so that the generation of the target vehicle attitude angle can simultaneously utilize the current height state and the height change trend at adjacent time moments. This allows the attitude adjustment in the protected subject to no longer rely solely on single-frame height feedback, but rather on vehicle attitude forward perception data oriented towards the continuous travel process.

[0033] Preferably, in the specific technical implementation of the step "assigning environmental obstacle weights to the terrain abrupt change features through the spatial attention layer in the attitude decision network to obtain the second hidden state representation", the technical essence of the spatial attention layer is that it is a data processing structure layer in the pre-configured attitude decision network used to filter out local height abrupt change regions with a high degree of influence on the target vehicle's attitude angle from the terrain abrupt change features. Its function is not to change the spatial meaning of the terrain abrupt change features themselves, but to assign different environmental obstacle weights to different local height abrupt change regions based on the spatial distribution density of abrupt change candidate points contained in the terrain abrupt change features, the outline geometry, and the positional relationship of the local height abrupt change regions in the preset global chassis coordinate system. The technical essence of the environmental obstacle weight assignment is to convert the influence of the local height abrupt change regions on the attitude adjustment of the all-terrain adaptive mobile chassis into weight data that can participate in the subsequent processing of the pre-configured attitude decision network. The weight data does not perform semantic classification of obstacle types, but rather adjusts the numerical contribution of different local height abrupt change regions in attitude adjustment based on the spatial distance between the local height abrupt change region and the grounding area of ​​multiple independently driven wheel leg composite units, the extension of the local height abrupt change region along the travel direction, and the spatial distribution density of abrupt change candidate points. When assigning environmental obstacle weights, the spatial attention layer first reads the spatial distribution density of candidate points in the terrain abrupt change features and forms a mutation concentration encoding data based on this density. This mutation concentration encoding data expresses the degree of concentration of elevation changes within a local height abrupt change area and subsequently participates in the formation of weight data. Next, the spatial attention layer reads the contour geometry in the terrain abrupt change features and extracts boundary extension direction encoding data based on this geometry. This boundary extension direction encoding data expresses the spatial orientation of the local height abrupt change area relative to the travel direction of the all-terrain adaptive mobile chassis. This boundary extension direction encoding data and the mutation concentration encoding data are jointly incorporated into the environmental obstacle weight assignment. Further, the spatial attention layer forms mutation area location encoding data based on the positional relationship of the local height abrupt change area in the preset global chassis coordinate system, and weights and correlates the mutation area location encoding data, mutation concentration encoding data, and boundary extension direction encoding data to generate a second hidden state representation.The technical essence of the second hidden state representation is that the terrain change features are weighted by environmental obstacles to form spatial risk-emphasis data. The spatial risk-emphasis data transforms the change concentration degree encoding data, boundary extension direction encoding data, and change area location encoding data of the local height change area into a spatial weight expression that can be further processed by the pre-configured attitude decision network. The second hidden state representation is then spliced ​​with the first hidden state representation to enable the generation process of the target vehicle attitude angle to distinguish between ordinary undulating areas and local height change areas that have a stronger impact on vehicle attitude.

[0034] Preferably, the specific implementation process of the step "feature splicing of the first hidden state representation and the second hidden state representation" is as follows: After the first hidden state representation has been formed by the state-aware coding layer and the second hidden state representation has been formed by the spatial attention layer, the first hidden state representation and the second hidden state representation are first processed to correspond fields, so that the order of the wheel-leg composite unit grounding area in the first hidden state representation and the order of the local height change area position in the second hidden state representation are kept in the same spatial caliber under the preset chassis global coordinate system. The reason for performing field correspondence processing is that the first hidden state representation mainly describes the height change relationship of the terrain elevation feedback sequence in the time direction, and the second hidden state representation mainly describes the influence relationship of the local height change area in the spatial direction of the terrain change feature. If the first hidden state representation and the second hidden state representation are directly input into the strategy output layer, the strategy output layer will find it difficult to distinguish between the continuous undulation change in the grounding area of ​​the same wheel-leg composite unit and the influence of the local height change in the grounding area of ​​adjacent wheel-leg composite units, which will easily cause the target vehicle attitude angle to be excessively affected by the height change of the grounding area of ​​a certain side wheel-leg composite unit. After completing the field mapping process, the first field sequence is formed by arranging the current height status and adjacent height change trends in the first hidden state representation using the wheel-leg composite unit installation node as an index. Then, the second field sequence is formed by arranging the mutation concentration encoding data, boundary extension direction encoding data, and mutation region location encoding data in the second hidden state representation using the wheel-leg composite unit grounding area associated with the same wheel-leg composite unit installation node as an index. Subsequently, the first and second field sequences are concatenated side-by-side according to the arrangement order of the wheel-leg composite unit installation nodes to form the concatenated feature. The first field sequence originates from the first hidden state representation and carries the current height status and adjacent height change trends. The second field sequence originates from the second hidden state representation and carries the mutation concentration encoding data, boundary extension direction encoding data, and mutation region location encoding data. The concatenation of the first and second field sequences according to the arrangement order of the wheel-leg composite unit installation nodes ensures that the concatenated feature retains the wheel-leg composite unit installation node correspondence between the first and second hidden state representations. The technical essence of the spliced ​​features is to merge the temporal height change information corresponding to the terrain elevation feedback sequence and the spatial abrupt change impact information corresponding to the terrain abrupt change features into the comprehensive attitude judgment data in the same input expression. The spliced ​​features are not simply connecting the two data results end to end, but rather, under the correspondence between the same wheel-leg composite unit installation node and the same wheel-leg composite unit grounding area, the temporal fluctuation information in the first hidden state representation can establish a field-level correspondence with the spatial abrupt change weight information in the second hidden state representation.This feature stitching method preserves both the representation of the terrain undulations and continuous change trends at the current moment by the first hidden state and the representation of the influence strength and contour geometry of the local height abrupt change regions by the second hidden state. The stitched features are then input into the strategy output layer, enabling the layer to read the overall terrain undulations, the morphology of local height abrupt change regions, and the influence relationships of the ground contact areas of each wheel leg composite unit within the same data structure. This provides a more complete data foundation for generating the target vehicle attitude angles.

[0035] Preferably, the specific implementation process of the step "inputting the stitched features into the strategy output layer of the attitude decision network to generate the target vehicle attitude angle for the all-terrain adaptive mobile chassis to maintain a baseline level state under the corresponding terrain" is as follows: After the stitched features are formed, the strategy output layer first reads the fields corresponding to the first hidden state representation and the second hidden state representation in the stitched features, and converts the stitched features into attitude angle candidate coding data according to the output field order in the network configuration record; the technical essence of the strategy output layer is a data processing structure layer in the pre-configured attitude decision network used to convert the stitched features into target vehicle attitude angles. It does not directly output fixed attitude values, but forms attitude angle candidate coding data that can be used to maintain a baseline level state based on the overall terrain undulation state, the morphology of local height change areas, and the influence relationship of the grounding areas of multiple independently driven wheel-leg composite units expressed by the stitched features. The attitude angle candidate coding data is used to represent the vehicle pitch adjustment direction and vehicle roll adjustment direction that the all-terrain adaptive mobile chassis can adopt under the corresponding terrain. The attitude angle candidate coding data is subsequently processed by attitude constraint filtering. During attitude constraint screening, the strategy output layer reads the baseline horizontal state of the all-terrain adaptive mobile chassis and performs deviation mapping processing between the baseline horizontal state and the candidate attitude angle encoding data to form attitude deviation mapping data. The technical essence of the baseline horizontal state is the vehicle reference attitude that the all-terrain adaptive mobile chassis is expected to maintain under the corresponding terrain. This baseline horizontal state participates in the formation of the attitude deviation mapping data, enabling it to express the direction and degree of deviation of the candidate attitude angle encoding data relative to the baseline horizontal state. Subsequently, the strategy output layer performs executability constraint processing on the attitude deviation mapping data based on the swing arm movement and telescopic movement range that multiple independently driven wheel-leg composite units can execute, to form executable attitude adjustment data. This executable attitude adjustment data is used to exclude attitude angle candidate encoding data that exceeds the adjustment capabilities of the wheel-leg composite units, allowing subsequent target vehicle attitude angles to continue entering the calculation process for swing arm adjustment angles and telescopic adjustment lengths. Finally, the strategy output layer generates the target vehicle attitude angle based on the executable attitude adjustment data. The technical essence of the target vehicle attitude angle is that the all-terrain adaptive mobile chassis needs to adopt the vehicle attitude control target to approach the reference level state under the combined effect of the current terrain elevation feedback sequence and the current terrain change characteristics. The target vehicle attitude angle is then used as the input for the adjustment parameter calculation, so that multiple independently driven wheel leg composite units can respectively calculate the corresponding swing arm adjustment angle and telescopic adjustment length, thereby forming a clear technical connection between the output result of the pre-configured attitude decision network and the subsequent execution action.

[0036] Preferably, the specific implementation process of the step "constructing training samples to train a pre-configured attitude decision network including a state-aware encoding layer, a spatial attention layer, and a policy output layer" is as follows: When the all-terrain adaptive mobile chassis enters the training data acquisition condition, the terrain elevation feedback sequence at multiple consecutive sampling moments is first synchronously acquired according to the structural installation topology of multiple independently driven wheel-leg composite units. At each sampling moment, the terrain abrupt change features obtained through spatial distribution analysis, the actual vehicle pitch angle, actual vehicle roll angle, swing arm adjustment angle, telescopic adjustment length, and chassis balance state features after performing the telescopic action are synchronously read. The technical essence of the training samples is that the attitude decision training record is formed by binding the terrain elevation feedback sequence, terrain abrupt change features, vehicle attitude change results, and chassis balance state features in the same process according to the same current moment. It is not an isolated stack of sensor readings, but a sample-carrying data that can express "what target vehicle attitude angle should be generated after the terrain input data enters the pre-configured attitude decision network". When constructing training samples, the terrain elevation feedback sequence from continuous sampling times is first written into the sample input field according to the arrangement order of the wheel-leg composite unit installation nodes. The spatial distribution density, contour geometry, and positional relationship of local height abrupt change regions from the terrain abrupt change features are also written into the same sample input field, ensuring that the training samples simultaneously encompass the overall terrain undulation state and the morphology of local height abrupt change regions. Subsequently, the target vehicle attitude angle is selected based on the chassis balance state characteristics after the extension / retraction action. This involves using historical vehicle attitude angles that have a low first deviation value for the current pitch angle from the horizontal plane and a low second deviation value for the current roll angle from the horizontal plane as the target vehicle attitude angle. A one-to-one correspondence is established between this target vehicle attitude angle and the corresponding terrain elevation feedback sequence and terrain abrupt change features. The technical essence of the target vehicle attitude angle is that it serves as supervisory data used to constrain the output direction of the policy output layer during training. The target vehicle attitude angle originates from the chassis balance state characteristics corresponding to the same training sample, rather than a fixed angle predetermined based on the results and detached from the terrain. The training samples formed in this way enable the pre-configured attitude decision network to learn the causal correspondence between terrain elevation feedback sequences, terrain change features, target vehicle attitude angles, and chassis balance state features during the training phase. Subsequent training processes can then adjust the network mapping parameters in the state perception encoding layer, spatial attention layer, and policy output layer around this causal correspondence.

[0037] Preferably, the state-aware coding layer adopts a one-dimensional temporal convolutional structure unfolded along the sampling time direction, and a gated state recursion structure is connected in series after the one-dimensional temporal convolutional structure. The technical essence of the one-dimensional temporal convolutional structure is to slide and read the local detection distance data change segments corresponding to the same wheel leg composite unit installation node on the terrain elevation feedback sequence of continuous sampling time, so as to extract the local change pattern of terrain undulation in the time direction. The technical essence of the gated state recursion structure is to selectively retain and attenuate the local change pattern extracted by the one-dimensional temporal convolutional structure, so that the terrain undulation change that has a continuous impact on the target vehicle attitude angle enters the first hidden state representation, while reducing the numerical contribution of short-term jitter or single ranging disturbance in the first hidden state representation. During the training of the state-aware encoding layer, terrain elevation feedback sequences at consecutive sampling times are first read from the training samples. The input fields of these sequences are then validated according to the network configuration records to form a training field-aligned terrain elevation feedback sequence. This training field-aligned terrain elevation feedback sequence is further arranged according to the wheel-leg composite unit installation nodes, ensuring that the local detection distance data of each wheel-leg composite unit installation node at different sampling times can enter the same temporal channel. Subsequently, a one-dimensional temporal convolutional structure is used to extract the training field-aligned terrain elevation feedback sequence, forming training local temporal fluctuation encoding data. This training local temporal fluctuation encoding data is used to express the direction and magnitude of height change of the grounding area of ​​the same wheel-leg composite unit in adjacent sampling times. The gated state recursive structure then recursively encodes the local temporal fluctuation encoding data used for training to form the first hidden state representation for training. This first hidden state representation establishes a training constraint relationship with the target vehicle attitude angles in the training samples, enabling the state-aware encoding layer to learn which continuous terrain height changes will cause the target vehicle attitude angles to change in the pitch direction and which will cause them to change in the roll direction during training. Through this training method, the network mapping parameters of the state-aware encoding layer do not merely fit the height values ​​at a single sampling moment, but rather learn the temporal evolution relationship of the terrain elevation feedback sequence over consecutive sampling moments. This allows the first hidden state representation to provide temporally continuous pre-perception data of vehicle attitude for subsequent feature stitching.

[0038] Preferably, the spatial attention layer adopts an attention weighting structure based on spatial adjacency relationships. These spatial adjacency relationships originate from a preset global chassis coordinate system, the structural installation topology of multiple independently driven wheel-leg composite units, and the positional relationships of local height abrupt change regions in the terrain abrupt change features. The technical essence of this spatial attention layer is not a typical two-dimensional image convolution structure, but rather a data processing structure layer that weights the numerical contributions of different local height abrupt change regions in the target vehicle attitude angle generation process based on the spatial position of the local height abrupt change region relative to the grounding area of ​​the wheel-leg composite unit, the spatial distribution density of abrupt change candidate points, and the contour geometry. When training the spatial attention layer, terrain abrupt change features are first read from the training samples, and the spatial distribution density of abrupt change candidate points in the terrain abrupt change features is normalized and encoded to form training abrupt change concentration encoding data; then, the boundary orientation of the contour geometry in the terrain abrupt change features is extracted to form training boundary extension direction encoding data; subsequently, based on the positional relationships of the local height abrupt change regions in the preset global chassis coordinate system, training abrupt change region position encoding data is formed. The training data on the concentration of mutations, the direction of boundary extension, and the location of mutation regions are collectively fed into the spatial attention layer. The spatial attention layer generates training environmental obstacle weights based on the spatial proximity between the training environmental obstacle location data and the grounding area of ​​the wheel-leg composite unit. These weights are then used to adjust the numerical contributions of the training environmental obstacle concentration data and the direction of boundary extension data, forming a representation of the second hidden state for training. To ensure that the training environmental obstacle weights reflect the impact of local height mutation regions on vehicle attitude changes, a spatial weight consistency loss function is introduced during training. This function constrains the training environmental obstacle weights to maintain consistency with the direction of change of chassis balance state characteristics in the training samples. Specifically, when the historical pass results corresponding to a certain local height mutation region cause a significant deviation in the current chassis pitch angle or roll angle, the spatial attention layer should assign a higher training environmental obstacle weight to that local height mutation region. Through this training method, the second hidden state representation formed by the spatial attention layer can transform the local height change region morphology in the terrain change feature into spatial risk-oriented data related to the target vehicle attitude angle, so that the second hidden state representation can form a complementary relationship with the first hidden state representation in subsequent feature stitching.

[0039] Preferably, during training, the first hidden state representation is formed by the state-aware coding layer according to the terrain elevation feedback sequence aligned with the training fields, and the second hidden state representation is formed by the spatial attention layer according to the training mutation concentration encoding data, the training boundary extension direction encoding data, and the training mutation region location encoding data. Subsequently, according to the arrangement order of the wheel-leg composite unit installation nodes, the first and second hidden state representations are processed to correspond to each other to form the training spliced ​​features. The training spliced ​​features enter the policy output layer, which adopts a multi-layer fully connected regression structure and an executable constraint projection structure. The technical essence of the multi-layer fully connected regression structure is to nonlinearly combine the temporal fluctuation information and spatial mutation weight information corresponding to the grounding areas of different wheel-leg composite units in the training spliced ​​features to form candidate encoding data for attitude angles. The technical essence of the executable constraint projection structure is to constrain the range of arm swinging and telescopic movements that multiple independently driven wheel-leg composite units can perform to the training attitude angle candidate encoding data to form executable attitude adjustment data for training. The executable attitude adjustment data used for training is further used to generate the target vehicle attitude angle for training, and compared with the sample target vehicle attitude angle in the training samples to adjust the network mapping parameters in the pre-configured attitude decision network. The loss function adopted is a comprehensive training loss function, which consists of an attitude angle deviation loss function, a baseline level deviation loss function, an action executability constraint loss function, a spatial weight consistency loss function, and a temporal smoothing loss function. Specifically, the attitude angle deviation loss function uses the mean squared difference between the training target vehicle attitude angle and the sample target vehicle attitude angle to constrain the attitude angle direction and amplitude output by the policy output layer; the baseline level deviation loss function uses the mean deviation between the predicted vehicle attitude corresponding to the training target vehicle attitude angle and the baseline level state to constrain the training target vehicle attitude angle to tend towards the baseline level state; the action executability constraint loss function penalizes whether the training executable attitude adjustment data exceeds the range of arm swing and extension / retraction movements to constrain the training target vehicle attitude angle to continue entering the calculation process of arm swing adjustment angle and extension / retraction length; the spatial weight consistency loss function constrains the correspondence between the weights of environmental obstacles and the characteristics of chassis balance; and the temporal smoothing loss function uses the difference in change between the training target vehicle attitude angle at adjacent sampling times to suppress sudden jumps in the training target vehicle attitude angle during continuous travel.By jointly training the state-aware encoding layer, spatial attention layer, and policy output layer using a comprehensive training loss function, the pre-configured attitude decision network can generate a target vehicle attitude angle that matches the corresponding terrain when both terrain elevation feedback sequence and terrain abrupt change features are input. This also enables the target vehicle attitude angle to form a continuous technical relationship with the subsequent calculation process of the swing arm adjustment angle and telescopic adjustment length.

[0040] Optionally, the step of calculating the swing arm adjustment angle and telescopic adjustment length corresponding to each of the plurality of independently driven wheel-leg composite units based on the target vehicle body attitude angle includes: Obtain the initial physical configuration parameters of the multiple independently driven wheel-leg composite units at the current moment; Construct the inverse kinematics transformation matrix of the all-terrain adaptive mobile chassis based on the target vehicle body attitude angle; Substitute the initial physical configuration parameters into the inverse kinematics transformation matrix to obtain the expected relative coordinate positions of each wheel-leg composite unit relative to the all-terrain adaptive mobile chassis. Based on the desired relative coordinate position, the expected rotation value of the rotary joint and the expected stroke value of the linear actuator corresponding to each of the wheel-leg composite units to reach the coordinate position are calculated respectively. Then, the expected rotation value of the rotary joint is used as the adjustment angle of the swing arm, and the expected stroke value of the linear actuator is used as the extension adjustment length.

[0041] Preferably, the specific implementation process of the step "obtaining the initial physical configuration parameters of the multiple independently driven wheel-leg composite units at the current moment" is as follows: After the target vehicle body attitude angle has been generated by the pre-configured attitude decision network, the configuration calibration record corresponding to the all-terrain adaptive mobile chassis is read first, and the execution status feedback record of the multiple independently driven wheel-leg composite units at the current moment is read simultaneously. The configuration calibration record is formed when the all-terrain adaptive mobile chassis completes the assembly calibration or starts the initialization calibration. The configuration calibration record includes the wheel-leg composite unit installation node corresponding to each of the multiple independently driven wheel-leg composite units, the installation offset of the wheel-leg composite unit installation node in the preset chassis global coordinate system, the rotary joint reference angle, the linear actuator reference length, the swing arm structure length, the wheel end grounding reference position, and the wheel-leg composite unit motion direction constraint. The execution status feedback record is derived from the current moment's rotary joint angle feedback, linear actuator length feedback, and wheel-end grounding status feedback. This record incorporates the mechanical states of multiple independently driven wheel-leg composite units at the current moment into the calculation, preventing subsequent calculations of the swing arm adjustment angle and telescopic adjustment length solely based on the static configuration fields in the configuration calibration record. Subsequently, the static configuration fields in the configuration calibration record are aligned with the current execution field in the execution status feedback record to form the initial physical configuration parameters. The static configuration field is derived from the configuration calibration record. This field provides the wheel-leg composite unit mounting node, mounting offset, rotary joint reference angle, linear actuator reference length, swing arm structure length, wheel-end grounding reference position, and wheel-leg composite unit motion direction constraints. The current execution field is derived from the execution status feedback record. This field provides the current rotary joint angle, current linear actuator length, and current wheel-end grounding reference position. After aligning the static configuration field and the current execution field according to the same wheel-leg composite unit mounting node, the initial physical configuration parameters can simultaneously express the assembly reference and current execution status of multiple independently driven wheel-leg composite units. The initial physical configuration parameters include the current rotary joint angle, current linear actuator length, current wheel-end grounding reference position, swing arm structure length, mounting offset, and wheel-leg composite unit motion direction constraints corresponding to each wheel-leg composite unit mounting node. These fields collectively describe the mechanical configuration of multiple independently driven wheel-leg composite units relative to the all-terrain adaptive mobile chassis at the current moment, rather than simply describing the angle of a single rotary joint or the length of a single linear actuator.The initial physical configuration parameters are used as input to the subsequent inverse kinematics transformation matrix because the target vehicle attitude angle only expresses the vehicle attitude control target that the all-terrain adaptive mobile chassis needs to approach, and does not directly give how each wheel-leg composite unit should swing or extend / retract. Only by combining the target vehicle attitude angle with the initial physical configuration parameters can the target vehicle attitude angle be converted into the desired relative coordinate position that multiple independently driven wheel-leg composite units can execute. Thus, the initial physical configuration parameters play a role in the entire control method, connecting the target vehicle attitude angle with the subsequent calculation process of swing arm adjustment angle and extension / retraction adjustment length. This ensures that the subsequent calculation is not an abstract calculation detached from the current mechanical state, but rather a unit-by-unit transformation based on the actual configuration of multiple independently driven wheel-leg composite units at the current moment.

[0042] Preferably, the specific implementation process of the step "constructing the inverse kinematic transformation matrix of the all-terrain adaptive mobile chassis based on the target vehicle attitude angle" is as follows: After the initial physical configuration parameters have been formed, the target pitch attitude component and the target roll attitude component contained in the target vehicle attitude angle are read first, and the preset chassis global coordinate system, the structural installation topology relationship of multiple independently driven wheel-leg composite units, and the installation offset in the configuration calibration record are read, thereby constructing the inverse kinematic transformation matrix of the all-terrain adaptive mobile chassis. The target pitch attitude component is derived from the target vehicle attitude angle, and the target pitch attitude component is used to express the vehicle attitude change that the all-terrain adaptive mobile chassis needs to adjust along the longitudinal coordinate direction; the target roll attitude component is derived from the target vehicle attitude angle, and the target roll attitude component is used to express the vehicle attitude change that the all-terrain adaptive mobile chassis needs to adjust along the lateral coordinate direction; the target pitch attitude component and the target roll attitude component are jointly entered into the construction process of the inverse kinematic transformation matrix, so that the inverse kinematic transformation matrix can simultaneously express the longitudinal attitude adjustment relationship and the lateral attitude adjustment relationship. The inverse kinematics transformation matrix is ​​not a general coordinate rotation table, but rather geometric transformation data used to describe the mapping relationship between the target vehicle body attitude angle and the desired relative coordinate positions of multiple independently driven wheel-leg composite units. When constructing the inverse kinematics transformation matrix, the vehicle body reference plane of the all-terrain adaptive mobile chassis is first used as the object to be adjusted. The longitudinal, lateral, and elevation coordinate directions in the preset chassis global coordinate system are used as a unified coordinate system. The target pitch attitude component is converted into an elevation compensation relationship along the longitudinal coordinate direction, and the target roll attitude component is converted into an elevation compensation relationship along the lateral coordinate direction. Subsequently, the installation offset of each wheel-leg composite unit mounting node relative to the chassis geometric center is written into the corresponding matrix field of the inverse kinematics transformation matrix, so that the influence of the target vehicle body attitude angle on different wheel-leg composite unit mounting nodes can be expanded according to the differences in installation position. The matrix field originates from the correspondence between the wheel-leg composite unit mounting node, installation offset, target pitch attitude component, and target roll attitude component. This matrix field subsequently participates in the coordinate transformation processing of the initial physical configuration parameters. The rows of the inverse kinematics transformation matrix correspond to the mounting nodes of each wheel-leg composite unit in multiple independently driven wheel-leg composite units. Each row represents the spatial transformation relationship required for a wheel-leg composite unit mounting node to transform from the current wheel-end grounding reference position to the desired relative coordinate position. The columns of the inverse kinematics transformation matrix correspond to the longitudinal coordinate components, lateral coordinate components, elevation coordinate components, and attitude compensation components related to the target vehicle body attitude angle in the preset chassis global coordinate system. The elements at the intersection of rows and columns represent the transformation coefficients of a certain wheel-leg composite unit mounting node on a certain coordinate component or attitude compensation component. These transformation coefficients are jointly determined by the target vehicle body attitude angle, the installation offset, and the structural installation topology.Through this construction method, the inverse kinematics transformation matrix can decompose the overall attitude target of the all-terrain adaptive mobile chassis into individual wheel-leg composite unit mounting nodes. This ensures that wheel-leg composite units at different locations do not respond to the same target vehicle body attitude angle with the same adjustment amount, but instead form different desired relative coordinate positions according to their position relative to the chassis geometric center. The inverse kinematics transformation matrix subsequently receives initial physical configuration parameters to convert the actual mechanical configuration at the current moment into desired relative coordinate positions that match the target vehicle body attitude angle, thereby providing a geometric calculation basis for the unit-by-unit solution of the swing arm adjustment angle and telescopic adjustment length.

[0043] Preferably, in the specific technical implementation of the step "substituting the initial physical configuration parameters into the inverse kinematics transformation matrix to obtain the desired relative coordinate position of each of the wheel-leg composite units relative to the all-terrain adaptive mobile chassis", the reason for substituting the initial physical configuration parameters into the inverse kinematics transformation matrix is ​​that the inverse kinematics transformation matrix only expresses the geometric transformation relationship between the target vehicle attitude angle and the mounting nodes of multiple wheel-leg composite units, while the initial physical configuration parameters express the current rotational joint angle, current linear actuator length, and current wheel-end grounding reference position of multiple independently driven wheel-leg composite units at the current moment; only by substituting the initial physical configuration parameters into the inverse kinematics transformation matrix can the target vehicle attitude angle that the all-terrain adaptive mobile chassis needs to adjust be converted into the desired relative coordinate position that each wheel-leg composite unit should move to relative to the all-terrain adaptive mobile chassis. In the specific substitution process, the initial physical configuration parameters are first organized into initial configuration coordinate data according to the arrangement order of the wheel-leg composite unit installation nodes. This initial configuration coordinate data includes the current wheel-end grounding reference position, current rotary joint angle, and current linear actuator length corresponding to each wheel-leg composite unit installation node. This initial configuration coordinate data originates from the initial physical configuration parameters and is used to convert the current mechanical state of multiple independently driven wheel-leg composite units into coordinate inputs that can be entered into the inverse kinematics transformation matrix. Subsequently, based on the matrix row in the inverse kinematics transformation matrix corresponding to the wheel-leg composite unit installation node, coordinate transformation processing is performed on the initial configuration coordinate data to obtain the desired relative coordinate position of the wheel-leg composite unit under the target vehicle body attitude angle constraint. This coordinate transformation processing, based on the transformation coefficients in the inverse kinematics transformation matrix, spatially maps the current wheel-end grounding reference position, current rotary joint angle, and current linear actuator length in the initial configuration coordinate data, transforming the initial configuration coordinate data from the current mechanical state into the desired relative coordinate position corresponding to the target vehicle body attitude angle. The desired relative coordinate positions include longitudinal relative coordinate positions, lateral relative coordinate positions, and elevation relative coordinate positions. The longitudinal relative coordinate position is used to express the relative position of the wheel-leg composite unit along the travel direction of the all-terrain adaptive mobile chassis. The lateral relative coordinate position is used to express the relative position of the wheel-leg composite unit in the left-right direction along the all-terrain adaptive mobile chassis. The elevation relative coordinate position is used to express the support height requirement of the wheel-leg composite unit relative to the vehicle body reference plane.For different wheel-leg composite units, the inverse kinematics transformation matrix generates different desired relative coordinate positions based on the installation offset of the wheel-leg composite unit mounting node and the target vehicle body attitude angle. For example, when the target vehicle body attitude angle requires the all-terrain adaptive moving chassis to reduce the front pitch deviation, the elevation relative coordinate position corresponding to the mounting node of the front wheel-leg composite unit will differ from the elevation relative coordinate position corresponding to the mounting node of the rear wheel-leg composite unit. This provides a positional basis for the subsequent swing arm and telescopic movements of different wheel-leg composite units. The role of the desired relative coordinate position in the protection subject is to convert the vehicle body attitude adjustment target into a ground position target that can be further solved by each wheel-leg composite unit. Subsequently, the expected value of the rotary joint rotation and the expected value of the linear actuator stroke are calculated based on the desired relative coordinate position, so that the swing arm adjustment angle and telescopic adjustment length no longer come from the preset operating mode, but from the coordinate transformation result formed by the target vehicle body attitude angle, initial physical configuration parameters, and inverse kinematics transformation matrix.

[0044] Preferably, the specific implementation process of the step "calculating the expected rotation value of the rotary joint and the expected stroke value of the linear actuator corresponding to the expected relative coordinate position of each wheel-leg composite unit, and then using the expected rotation value of the rotary joint as the adjustment angle of the swing arm and the expected stroke value of the linear actuator as the telescopic adjustment length" is as follows: After the expected relative coordinate positions of each wheel-leg composite unit have been formed, for each wheel-leg composite unit, the current rotary joint angle, current linear actuator length, swing arm structure length, and current wheel end grounding reference position corresponding to the wheel-leg composite unit are first read, and these parameters are geometrically correlated with the expected relative coordinate position of the wheel-leg composite unit to form joint stroke solution data. The joint stroke solution data comes from the geometric correspondence processing between the current rotary joint angle, current linear actuator length, swing arm structure length, current wheel end grounding reference position, and expected relative coordinate position. The joint stroke solution data is used to express the spatial difference between the current wheel end grounding reference position and the expected relative coordinate position. The joint stroke solution data is subsequently used in the rotary joint rotation expectation value calculation process and the linear actuator stroke expectation value calculation process, respectively. When calculating the expected rotation value of a rotary joint, the target swing arm direction to which the wheel-leg composite unit needs to swing is first determined based on the longitudinal and lateral relative coordinate positions in the desired relative coordinate position. Then, combined with the swing arm structure length and the current rotary joint angle, the change in rotary joint angle required to move the current wheel-end grounding reference position along the target swing arm direction is calculated to obtain the expected rotation value of the rotary joint. The target swing arm direction is derived from the longitudinal and lateral relative coordinate positions and is used to define the directional calculation caliber of the expected rotation value of the rotary joint. The change in rotary joint angle is derived from the target swing arm direction, the swing arm structure length, and the current rotary joint angle, and is used to form the expected rotation value of the rotary joint. The expected rotation value of the rotary joint describes the angle change required for the rotary joint to rotate from the current rotary joint angle to meet the desired relative coordinate position. It is subsequently used as the swing arm adjustment angle, enabling the corresponding wheel-leg composite unit to adjust the horizontal support orientation of the wheel end through swing arm movement. When calculating the expected value of the linear actuator stroke, first determine the support height change that the wheel-leg composite unit needs to compensate for based on the elevation relative coordinate position in the expected relative coordinate position and the current length of the linear actuator. Then, combine the current wheel end grounding reference position and the target swing arm direction to calculate the amount of change in the linear actuator length required for the linear actuator to extend or retract from the current linear actuator length to meet the expected relative coordinate position, so as to obtain the expected value of the linear actuator stroke.The change in support height originates from the correspondence between the relative elevation coordinate position and the current length of the linear actuator. This change in support height is used to define the length calculation caliber for the expected stroke value of the linear actuator. The change in linear actuator length originates from the change in support height, the current wheel-end grounding reference position, and the target swing arm direction. This change in linear actuator length is used to form the expected stroke value of the linear actuator. The expected stroke value of the linear actuator describes the length change required for the linear actuator to adjust from its current length to meet the desired relative coordinate position. This is subsequently used as the telescopic adjustment length, enabling the corresponding wheel-leg composite unit to adjust the support height of the wheel end through telescopic movements after the swing arm action is completed. By using the expected rotation value of the rotary joint as the swing arm adjustment angle and the expected stroke value of the linear actuator as the telescopic adjustment length, the target vehicle body attitude angle, after a step-by-step transformation using the inverse kinematics transformation matrix and the desired relative coordinate position, can be converted into the swing arm action and telescopic action parameters executable by each wheel-leg composite unit. This provides a continuous technical link from the target attitude to the execution parameters for the attitude adjustment of the all-terrain adaptive mobile chassis under abrupt terrain changes.

[0045] Optionally, the step of calculating the expected rotation value of the rotary joint and the expected stroke value of the linear actuator corresponding to the expected relative coordinate position for each of the wheel-leg composite units includes: For each of the wheel-leg composite units, the corresponding grounding contact position is extracted based on the desired relative coordinate position to establish a first kinematic geometric constraint condition that reflects the relationship between the swing arm adjustment angle and the corresponding grounding contact position. Establish a second kinematic geometric constraint condition that reflects the relationship between the telescopic adjustment length and the position of the grounding contact; Under the premise of satisfying the first kinematic geometric constraints and the second kinematic geometric constraints, the optimization objective is to minimize the overall energy consumed by the wheel-leg composite unit in performing the swing arm movement and the extension movement. Then, within the feasible domain of the joint, the expected value of the rotation of the rotary joint and the expected value of the stroke of the linear actuator are calculated.

[0046] Preferably, the specific implementation process of step "for each of the wheel-leg composite units, extract the corresponding grounding contact position based on the desired relative coordinate position" is as follows: After the desired relative coordinate position of each wheel-leg composite unit has been generated by the inverse kinematic transformation matrix, the desired relative coordinate position corresponding to each wheel-leg composite unit is read according to the arrangement order of the wheel-leg composite unit installation nodes, and the current wheel-end grounding reference position, swing arm structure length, current rotary joint angle, current linear actuator length, and wheel-leg composite unit motion direction constraint corresponding to the wheel-leg composite unit are read simultaneously. The desired relative coordinate position includes longitudinal relative coordinate position, lateral relative coordinate position, and elevation relative coordinate position. The longitudinal relative coordinate position is used to limit the target relative position of the wheel-leg composite unit along the travel direction of the all-terrain adaptive mobile chassis, the lateral relative coordinate position is used to limit the target relative position of the wheel-leg composite unit in the left and right directions along the all-terrain adaptive mobile chassis, and the elevation relative coordinate position is used to limit the target support height of the wheel-leg composite unit relative to the vehicle body reference plane. When extracting the grounding contact point location, the target horizontal projection position of the wheel end in the preset chassis global coordinate system is first determined based on the longitudinal and lateral relative coordinate positions. Then, the target support height corresponding to the target horizontal projection position is determined based on the elevation relative coordinate position. The target horizontal projection position and the target support height are combined to form the grounding contact point location. The grounding contact point location is not simply the geometric center position of the wheel end, but rather the spatial target position where the wheel-leg composite unit should form a wheel-end support relationship with the ground under the constraint of the target vehicle body attitude angle. The wheel-end support relationship represents the spatial correspondence of the support formed at the grounding contact point position after the wheel-leg composite unit adjusts through swing arm movement and extension / retraction movement. This wheel-end support relationship is subsequently defined by both the first and second kinematic geometric constraints. This grounding contact point location not only inherits the spatial target in the desired relative coordinate position but also provides a common constraint benchmark for subsequently establishing the first and second kinematic geometric constraints. The reason for this approach is that the swing arm adjustment angle primarily changes the horizontal support orientation at the wheel end, while the telescopic adjustment length primarily changes the support height at the wheel end. Both must be calculated around the same ground contact point to avoid inconsistencies in the wheel end support relationship caused by the swing arm movement and telescopic movement pointing to different target positions. Therefore, the ground contact point position transforms the desired relative coordinate position into a spatial target position that the wheel-leg composite unit can recognize, allowing the target at the vehicle body attitude level to be further incorporated into the calculation of the expected rotation values ​​of the rotary joints and the expected stroke values ​​of the linear actuators.

[0047] Preferably, in the specific technical implementation of the step "establishing a first kinematic geometric constraint condition reflecting the relationship between the swing arm adjustment angle and the corresponding ground contact position", after the ground contact position has been determined, the center position of the rotary joint corresponding to the wheel-leg composite unit, the length of the swing arm structure, the current rotary joint angle, and the motion direction constraint of the wheel-leg composite unit are first read. The center position of the rotary joint is derived from the joint installation position corresponding to the installation node of the wheel-leg composite unit in the configuration calibration record. The center position of the rotary joint is used to represent the geometric reference around which the swing arm movement generates angular changes. The center position of the rotary joint and the ground contact position subsequently participate in the formation of the horizontal projection data of the swing arm. Subsequently, the center position of the rotary joint and the ground contact position are horizontally projected and correspondingly processed in a preset chassis global coordinate system to form the horizontal projection data of the swing arm. The horizontal projection data of the swing arm is used to express the relative offset relationship between the center position of the rotary joint and the ground contact position in the longitudinal and lateral coordinate directions. The horizontal projection data of the swing arm subsequently participates in the solution of the swing arm adjustment angle. When establishing the first kinematic geometric constraint, the horizontal position range that the wheel-leg composite unit can cover through the swing arm movement is first determined based on the swing arm structure length and horizontal projection data. Then, the allowable angle range of the rotary joint is determined based on the current rotary joint angle and the motion direction constraint of the wheel-leg composite unit. Subsequently, the grounding contact position is limited to the horizontal position range that the swing arm movement can cover, thus forming the first kinematic geometric constraint. The allowable angle range of the rotary joint is used to represent the angular change boundary that the rotary joint can perform based on the current rotary joint angle. The allowable angle range of the rotary joint is subsequently used to construct the feasible domain range of the joint. The first kinematic geometric constraint is used to describe the geometric correspondence between the swing arm adjustment angle, the rotary joint center position, the swing arm structure length, and the grounding contact position. Its function is to convert the horizontal position requirement in the grounding contact position into the angular change requirement that the rotary joint can perform. The reason for establishing the first kinematic geometric constraint is that the swing arm movement of the wheel-leg composite unit is not a translation of the wheel end in any arbitrary direction, but rather an angular change around the center position of the rotary joint. If the swing arm adjustment angle is directly generated based solely on the ground contact point position, the length of the swing arm structure and the allowable angle range of the rotary joint are easily overlooked, making the calculated swing arm adjustment angle unexecutable by the wheel-leg composite unit. Through the first kinematic geometric constraint, a calculable geometric relationship is formed between the swing arm adjustment angle and the ground contact point position. The subsequent calculation of the expected value of the rotary joint rotation can be performed within the allowable angle range of the rotary joint, enabling the swing arm adjustment angle to serve the horizontal support orientation adjustment of the wheel end at the ground contact point position, and connecting with the goal of maintaining a baseline level state of the all-terrain adaptive mobile chassis under abrupt terrain changes.

[0048] Preferably, the specific implementation process of the step "establishing a second kinematic geometric constraint condition reflecting the relationship between the telescopic adjustment length and the grounding contact position" is as follows: After the first kinematic geometric constraint condition has been formed, the current linear actuator length, linear actuator reference length, linear actuator adjustable stroke range, current wheel end grounding reference position, and grounding contact position corresponding to the same wheel leg composite unit are read. The height difference is extracted based on the elevation relative coordinate position in the grounding contact position and the current support height in the current wheel end grounding reference position to form support height compensation data. The current support height originates from the elevation coordinate position of the current wheel end grounding reference position in a preset chassis global coordinate system. The current support height is used to extract the height difference with the elevation relative coordinate position. The support height compensation data is used to express the support height change that needs to be compensated when the wheel leg composite unit reaches the grounding contact position from the current wheel end grounding reference position. The support height compensation data subsequently participates in the establishment of the second kinematic geometric constraint condition. When establishing the second kinematic geometric constraint, the horizontal support orientation of the wheel end after the swing arm movement is first determined based on the horizontal projection data of the swing arm obtained from the first kinematic geometric constraint. Then, the required extension or retraction length range of the linear actuator is determined based on the support height compensation data and the current length of the linear actuator. Subsequently, a correspondence is established between the adjustable stroke range of the linear actuator and the relative elevation coordinates of the grounding contact position to form the second kinematic geometric constraint. The second kinematic geometric constraint describes the geometric correspondence between the telescopic adjustment length, the current length of the linear actuator, the support height compensation data, and the grounding contact position. Its function is to convert the elevation position requirement in the grounding contact position into the length change requirement that the linear actuator can perform. The reason for establishing the second kinematic geometric constraint is that the telescopic movement of the wheel-leg composite unit typically changes the wheel-end support height along the direction of motion of the linear actuator. This change in wheel-end support height is influenced by the adjustable stroke range of the linear actuator, the horizontal support orientation of the wheel end after the swing arm movement, and the current grounding reference position of the wheel end. Without the second kinematic geometric constraint, the telescopic adjustment length might only meet the height change requirement but not the horizontal support orientation of the wheel end after the swing arm movement. Through the second kinematic geometric constraint, the telescopic adjustment length can be solved around the same grounding contact point, and together with the first kinematic geometric constraint, it defines the executable motion relationship of the wheel-leg composite unit, ensuring that the subsequent expected stroke value of the linear actuator and the expected rotation value of the rotary joint are formed collaboratively at the same grounding contact point position.

[0049] Preferably, the specific implementation process of the step "under the premise of satisfying the first kinematic geometric constraints and the second kinematic geometric constraints, with the optimization objective of minimizing the overall energy consumed by the wheel-leg composite unit in performing the swing arm movement and the telescopic movement, and then calculating the expected value of the rotation of the rotary joint and the expected value of the stroke of the linear actuator within the joint feasible domain" is as follows: After the first and second kinematic geometric constraints have been established, the joint feasible domain is constructed based on the current rotation joint angle, the allowable angle range of the rotation joint, the current length of the linear actuator, the adjustable stroke range of the linear actuator, the motion direction constraint of the wheel-leg composite unit, and the position of the grounding contact. The joint feasible domain is used to represent the candidate rotation joint rotation value and the candidate linear actuator stroke value that can be adopted by the rotary joint and the linear actuator without deviating from the first kinematic geometric constraints, the second kinematic geometric constraints, the allowable angle range of the rotation joint, and the adjustable stroke range of the linear actuator. Subsequently, multiple candidate joint stroke combinations are generated within the feasible joint domain. Each candidate joint stroke combination includes candidate rotary joint rotation values ​​and candidate linear actuator stroke values. For each candidate joint stroke combination, the rotation values ​​are first verified according to the first kinematic geometric constraint to ensure that the horizontal support orientation of the wheel end is close to the ground contact point. Then, the linear actuator stroke values ​​are verified according to the second kinematic geometric constraint to ensure that the height of the wheel end support is close to the ground contact point. Candidate joint stroke combinations that cannot simultaneously satisfy the first and second kinematic geometric constraints are eliminated. For the remaining candidate joint stroke combinations, the overall energy index is further calculated. The overall energy index is used to evaluate the angle change range, length change range, and coordination degree between the swing arm action and the telescopic action required by the wheel-leg composite unit when performing the swing arm action and the telescopic action. It does not evaluate the swing arm action or the telescopic action separately, but rather incorporates the angle change burden corresponding to the candidate rotary joint rotation value and the length change burden corresponding to the candidate linear actuator stroke value into the same evaluation caliber. The angular change burden originates from the variation in the candidate rotary joint rotation value relative to the current rotary joint angle, and the length change burden originates from the variation in the candidate linear actuator stroke value relative to the current linear actuator length. Both the angular change burden and the length change burden participate in the calculation of the overall energy index. The reason for minimizing the overall energy index as the optimization objective is that, under abrupt terrain changes, multiple candidate joint stroke combinations can all enable the wheel-leg composite unit to reach the ground contact position, but the swing arm movement amplitude and extension / retraction movement amplitude corresponding to different candidate joint stroke combinations are different. Selecting candidate joint stroke combinations with a lower overall energy index can reduce unnecessary large swing arm movements or large extension / retraction movements, making the execution process of the wheel-leg composite unit more aligned with the current vehicle body posture adjustment requirements.Finally, the candidate rotary joint rotation value with the lowest overall energy index that satisfies both the first and second kinematic geometric constraints is determined as the expected rotation value of the rotary joint. Similarly, the candidate linear actuator stroke value within the same candidate joint stroke combination is determined as the expected linear actuator stroke value. The expected rotation value of the rotary joint is subsequently used as the swing arm adjustment angle, and the expected linear actuator stroke value is subsequently used as the telescopic adjustment length. This allows each wheel-leg composite unit to sequentially perform swing arm and telescopic movements under the constraint of the same ground contact position, thereby converting the desired relative coordinate position into joint adjustment parameters that can be executed by the wheel-leg composite unit.

[0050] Optionally, the step of controlling the corresponding wheel-leg composite unit to perform the swing arm movement according to the swing arm adjustment angle, and controlling the corresponding wheel-leg composite unit to perform the telescopic movement according to the telescopic adjustment length after the swing arm movement is completed, includes: A first drive command stream is generated based on the swing arm adjustment angle to instruct the swing arm movement, and then the first drive command stream is sent to the rotary joint actuator corresponding to the wheel-leg composite unit to control it to perform the swing arm movement. The angle encoder within the wheel-leg composite unit monitors the swing arm's execution angle in real time. When it is determined that the swing arm execution angle has reached the swing arm adjustment angle and the swing arm action has been completed, a second drive command stream is generated based on the telescopic adjustment length to indicate the telescopic action. The second drive command stream is then sent to the linear drive actuator corresponding to the wheel-leg composite unit to control it to execute the telescopic action.

[0051] Preferably, the specific implementation process of step "generating a first drive command stream to indicate the swing arm movement based on the swing arm adjustment angle" is as follows: After the swing arm adjustment angle has been determined by the rotational joint rotation expectation value, the current rotational joint angle, the allowable angle range of the rotational joint, the motion direction constraint of the wheel-leg composite unit, the command receiving format of the rotational joint actuator, and the current control cycle flag corresponding to the wheel-leg composite unit are read first, and the above data and the swing arm adjustment angle are written into the swing arm command encapsulation record. The command receiving format of the rotational joint actuator is configured in the configuration calibration record corresponding to the rotational joint actuator identifier. The command receiving format of the rotational joint actuator is used to limit the field order, field length, and verification method in the swing arm command segment. The command receiving format of the rotational joint actuator subsequently participates in the segmented encoding of the swing arm adjustment angle. The arm control instruction encapsulation record is used to incorporate the source of the arm control adjustment angle, the instruction receiving format of the rotary joint actuator, the current rotary joint angle, the allowable angle range of the rotary joint, and the current control cycle marker into the same instruction generation caliber. This ensures that the first drive instruction stream corresponds to the arm control action of the current wheel-leg composite unit, rather than being misused in other control cycles or other wheel-leg composite units. Subsequently, the arm control action direction is determined based on the angular difference between the current rotary joint angle and the arm control adjustment angle, and the arm control action direction is boundary-checked based on the allowable angle range of the rotary joint to form an arm control angle check result. The arm control angle check result continues to participate in the generation of the first drive instruction stream, ensuring that the first drive instruction stream only contains arm control action information within the allowable angle range of the rotary joint. Subsequently, the arm control adjustment angle is segmented and encoded according to the control beat corresponding to the current control cycle marker and the instruction receiving format of the rotary joint actuator to form multiple arm control instruction segments. Each arm control instruction segment includes the wheel-leg composite unit installation node, the rotary joint actuator identifier, the arm control action direction, the segmented arm control angle, the segmented execution sequence, and the instruction check marker. The arm control command fragment is not a single angle value, but rather a sequence of drive data fragments that the rotary joint actuator can read sequentially. These arm control command fragments are concatenated in a segmented execution sequence to form a first drive command stream. This first drive command stream instructs the rotary joint actuator to perform arm movements according to the arm adjustment angle. In the control method of this application, it inherits the arm adjustment angle and provides a common arm adjustment angle reference for subsequent real-time monitoring of the arm execution angle. By first forming an arm control command encapsulation record, and then generating arm angle verification results and arm control command fragments from the encapsulation record, thus forming the first drive command stream, the arm adjustment angle can be converted from a calculated result into ordered drive data that the rotary joint actuator can execute.

[0052] Preferably, in the specific technical implementation of the step "sending the first drive command stream to the rotary joint actuator corresponding to the wheel-leg composite unit to control it to execute the swing arm movement", after the first drive command stream has been formed, the wheel-leg composite unit mounting node and the rotary joint actuator identifier in the first drive command stream are first read, and the corresponding rotary joint actuator is determined from the configuration calibration record based on the wheel-leg composite unit mounting node, so that the first drive command stream can be sent to the rotary joint actuator corresponding to the wheel-leg composite unit. The rotary joint actuator is used to receive the first drive command stream and drive the wheel-leg composite unit to generate angular changes around the center position of the rotary joint according to the swing arm movement direction, segmented swing arm angle and segmented execution sequence in the first drive command stream. When sending the first drive command stream, it is first written into the swing arm action distribution record. The swing arm action distribution record includes the wheel-leg composite unit installation node, the rotary joint actuator identifier, the first drive command stream, the current control cycle marker, and the swing arm adjustment angle. The swing arm action distribution record is used to uniformly record the recipient of the first drive command stream, the current control cycle marker, and the swing arm adjustment angle. Subsequently, according to the rotary joint actuator identifier in the swing arm action distribution record, the first drive command stream is sent to the rotary joint actuator. After receiving the first drive command stream, the rotary joint actuator first verifies the command verification marker in the first drive command stream. After the verification is passed, it reads the swing arm command segments sequentially according to the segmented execution order and executes the swing arm action according to the segmented swing arm angles in the swing arm command segments, so that the horizontal support position of the wheel end of the wheel-leg composite unit gradually approaches the horizontal support position of the wheel end corresponding to the swing arm adjustment angle. The reason for controlling the wheel-leg composite unit to execute the swing arm movement first is that the desired relative coordinate position has been decomposed into two interrelated adjustment requirements: the wheel-end horizontal support orientation and the wheel-end support height. The swing arm movement mainly addresses the wheel-end horizontal support orientation adjustment problem. If the telescopic movement is executed directly before the swing arm movement is completed, the support height compensation corresponding to the telescopic adjustment length may be applied based on the wheel-end horizontal support orientation that has not yet been established, resulting in a deviation between the grounding contact position and the wheel-end support relationship. By first driving the rotary joint actuator to execute the swing arm movement through the first drive command stream, the wheel-leg composite unit can first establish a wheel-end horizontal support orientation that matches the grounding contact position, providing a prerequisite position condition for subsequent telescopic movements according to the telescopic adjustment length.

[0053] Preferably, the specific implementation process of step "real-time monitoring of the swing arm execution angle through the angle encoder in the wheel-leg composite unit" is as follows: During the process of the rotary joint actuator performing the swing arm action according to the first drive command stream, the angle encoder reads the angle feedback data of the rotary joint according to the sampling beat corresponding to the current control cycle mark, and converts the angle feedback data of the rotary joint into the swing arm execution angle. The angle encoder is set in the wheel-leg composite unit and corresponds to the angle change position of the rotary joint. The angle encoder is used to convert the actual angle change of the rotary joint during the swing arm action into the swing arm execution angle that can be compared with the swing arm adjustment angle in the same dimension. When monitoring the swing arm execution angle in real time, the swing arm adjustment angle and the current control cycle mark in the swing arm action issuance record are read first, and then the angle feedback data of the rotary joint returned by the angle encoder under the same current control cycle mark is read. The angle feedback data of the rotary joint is zero-point reference correction to form the swing arm execution angle. The zero-point reference correction is based on the rotary joint reference angle in the configuration calibration record. The rotary joint reference angle is used to eliminate the initial angle difference of different wheel-leg composite units in the assembly position, so that the swing arm execution angle can correspond to the same rotary joint angle caliber. Subsequently, the swing arm execution angle is written into the swing arm angle monitoring record. This record includes the wheel-leg composite unit installation node, the rotary joint actuator identifier, the swing arm adjustment angle, the swing arm execution angle, the current control cycle marker, and angle difference data. The angle difference data is generated by comparing the swing arm execution angle and the swing arm adjustment angle using the same dimensions. This angle difference data is subsequently used to determine whether the swing arm execution angle has reached the swing arm adjustment angle. The reason for real-time monitoring of the swing arm execution angle using an angle encoder is that the first drive command stream can only express the desired swing arm movement and cannot directly reflect whether the rotary joint actuator has adjusted the wheel-leg composite unit to the swing arm adjustment angle under abrupt terrain changes. When the wheel-leg composite unit is affected by ground reaction forces, structural clearances, or load changes, the actual angle of the rotary joint actuator may deviate from the swing arm adjustment angle corresponding to the first drive command stream. Therefore, the swing arm execution angle plays a role in connecting the swing arm action execution process and the extension and retraction action triggering conditions in the control method of this application. Only through continuous comparison between the swing arm execution angle and the swing arm adjustment angle can the sequential control relationship of "completing the swing arm action and then performing the extension and retraction action" be implemented as a judgmentable technical condition.

[0054] Preferably, the specific implementation process of the step "when it is determined that the swing arm execution angle has reached the swing arm adjustment angle, and the swing arm action is determined to be completed, a second drive command stream for indicating the telescopic action is generated according to the telescopic adjustment length, and then the second drive command stream is sent to the linear drive actuator corresponding to the wheel-leg composite unit to control it to execute the telescopic action" is as follows: After the swing arm angle monitoring record has been formed, the swing arm execution angle, swing arm adjustment angle and angle difference data in the swing arm angle monitoring record are read first, and the angle difference data is compared with the preset angle arrival judgment range; the preset angle arrival judgment range is configured in the configuration calibration record according to the measurement resolution capability of the angle encoder, the angle response capability of the rotary joint actuator and the motion direction constraint of the wheel-leg composite unit, and the preset angle arrival judgment range is used to determine whether the swing arm execution angle has reached the swing arm adjustment angle. When the angle difference data falls within the preset angle positioning range, it is determined that the swing arm execution angle has reached the swing arm adjustment angle, and a swing arm action completion marker is written into the swing arm angle monitoring record. This swing arm action completion marker originates from the comparison result between the swing arm execution angle and the swing arm adjustment angle, and subsequently serves as the trigger condition for generating the second drive command stream. Subsequently, the telescopic adjustment length, current linear actuator length, adjustable stroke range of the linear actuator, wheel-leg composite unit motion direction constraint, command receiving format of the linear drive actuator, and the swing arm action completion marker are read, and the above data is written into the telescopic command encapsulation record. The command receiving format of the linear drive actuator is configured in the configuration calibration record corresponding to the linear drive actuator identifier. The command receiving format of the linear drive actuator is used to limit the field order, field length, and verification method in the telescopic command segment, and subsequently participates in the segmented encoding of the telescopic adjustment length. The telescopic command encapsulation record is used to incorporate the source of the telescopic adjustment length, the command receiving format of the linear drive actuator, the current length of the linear actuator, the adjustable stroke range of the linear actuator, and the completion mark of the swing arm movement into the same command generation caliber, ensuring that the second drive command stream can only be formed after the swing arm movement is completed. When generating the second drive command stream based on the telescopic command encapsulation record, the telescopic movement direction is first determined based on the length difference between the current linear actuator length and the telescopic adjustment length. Then, the telescopic movement direction is boundary-checked based on the adjustable stroke range of the linear actuator to form a telescopic length check result. The telescopic length check result continues to participate in the generation of the second drive command stream, ensuring that the second drive command stream only contains telescopic movement information within the adjustable stroke range of the linear actuator.Subsequently, the telescopic adjustment length is segmented and encoded according to the control cycle marker corresponding to the current control cycle and the instruction receiving format of the linear drive actuator to form multiple telescopic instruction fragments. These fragments are then concatenated into a second drive instruction stream according to their segment execution order. Each telescopic instruction fragment includes a wheel-leg composite unit mounting node, a linear drive actuator identifier, a telescopic action direction, segmented telescopic length, segment execution order, and an instruction verification flag. The second drive instruction stream instructs the linear drive actuator to perform the telescopic action according to the telescopic adjustment length. The linear drive actuator receives the second drive instruction stream and drives the wheel-leg composite unit to change the wheel end support height based on the telescopic action direction, segmented telescopic length, and segment execution order in the second drive instruction stream. Finally, the second drive instruction stream is written into the telescopic action distribution record and sent to the linear drive actuator according to the linear drive actuator identifier in the record to control the linear drive actuator to perform the telescopic action. By using the swing arm movement to complete the generation of the second drive command stream, the telescopic movement can be established on the already positioned wheel end horizontal support orientation, thus creating a clear sequential relationship between the wheel end support height adjustment corresponding to the telescopic adjustment length and the wheel end horizontal support orientation adjustment corresponding to the preceding swing arm adjustment angle.

[0055] Optionally, the iterative update of the network mapping parameters in the attitude decision network based on the chassis balance state characteristics after performing the above-mentioned telescopic action includes: The inertial measurement unit installed on the all-terrain adaptive mobile chassis collects the current pitch angle and current roll angle of the chassis after performing the above-mentioned telescopic action. Based on the first deviation value of the current pitch angle of the chassis from the horizontal plane and the second deviation value of the current roll angle of the chassis from the horizontal plane, an environmental interaction reward value is constructed, and then the environmental interaction reward value is used as the chassis balance state feature. For the attitude decision network, a time difference error is calculated based on the environmental interaction reward value, which is a feature of the chassis balance state, and then gradient information is calculated based on the time difference error. The network mapping parameters in the attitude decision network are iteratively updated according to the descent direction corresponding to the gradient information in order to maximize the cumulative environmental interaction reward value generated by subsequent actions.

[0056] Preferably, the specific implementation process of the step "collecting the current pitch angle and current roll angle of the chassis after performing the above-mentioned telescopic action by means of an inertial measurement unit set on the all-terrain adaptive mobile chassis" is as follows: After the corresponding wheel-leg composite unit has performed the telescopic action according to the telescopic adjustment length, the installation node of the wheel-leg composite unit, the telescopic adjustment length, the current control cycle mark, and the telescopic action completion status in the telescopic action distribution record are read first, and the above data are written into the attitude acquisition trigger record, so that the current pitch angle and current roll angle of the chassis obtained subsequently can correspond to the chassis balance state after this swing arm action and telescopic action. The inertial measurement unit is set on the vehicle body reference plane of the all-terrain adaptive mobile chassis. The inertial measurement unit is used to collect angular velocity feedback data and acceleration feedback data of the vehicle body reference plane relative to the horizontal plane, and through attitude fusion processing of angular velocity feedback data and acceleration feedback data, the current pitch angle and current roll angle of the chassis that can characterize the tilt state of the vehicle body reference plane are formed. The attitude acquisition trigger record is also used to record the attitude stabilization sampling time window corresponding to the current control cycle mark. This attitude stabilization sampling time window is configured in the configuration calibration record based on the completion state of the telescopic movement, the sampling cycle of the inertial measurement unit, and the vehicle body vibration attenuation requirements after the wheel-leg composite unit performs the telescopic movement. The attitude stabilization sampling time window ensures that the current pitch angle and roll angle of the chassis originate from the stable attitude feedback after the telescopic movement is completed, rather than from the transient attitude feedback during the telescopic movement. During acquisition, the inertial measurement unit is first triggered according to the current control cycle mark to continuously read angular velocity feedback data and acceleration feedback data within the attitude stabilization sampling time window. Then, the angular velocity feedback data is time-integrated to form angle change estimation data, and the acceleration feedback data is decomposed by gravity direction to form horizontal reference direction data. Subsequently, the angle change estimation data and the horizontal reference direction data are attitude-fused to obtain the current pitch angle and roll angle of the chassis. The current pitch angle of the chassis represents the longitudinal tilt of the all-terrain adaptive mobile chassis relative to the horizontal plane, and the current roll angle represents the lateral tilt of the all-terrain adaptive mobile chassis relative to the horizontal plane. Both together reflect the deviation of the vehicle's reference plane from the baseline horizontal state after performing the aforementioned telescopic maneuver. Subsequently, the current pitch angle, current roll angle, current control cycle marker, attitude stability sampling time window, and the corresponding telescopic maneuver completion status are written into the chassis attitude feedback record. This chassis attitude feedback record subsequently participates in the formation of the first deviation value, the second deviation value, and the environmental interaction reward value, ensuring that the chassis balance state characteristics do not solely depend on the command issuance result but originate from the actual attitude feedback after performing the aforementioned telescopic maneuver.In this way, the current pitch angle and current roll angle of the chassis collected by the inertial measurement unit can introduce the chassis balance state after the wheel leg composite unit performs the swing arm action and extension action into the iterative update process of the attitude decision network, so that the subsequent generation of the target vehicle attitude angle can absorb the actual attitude response of the previous control cycle.

[0057] Preferably, in the specific technical implementation of the step "for the attitude decision network, calculate the time difference error based on the environmental interaction reward value as a feature of the chassis balance state, and then calculate the gradient information based on the time difference error", after the chassis attitude feedback record has been formed, the current pitch angle, current roll angle, and reference horizontal state of the chassis are first read, and the first deviation value of the current pitch angle from the horizontal plane and the second deviation value of the current roll angle from the horizontal plane are calculated respectively; then, the environmental interaction reward value is constructed based on the first and second deviation values, and the environmental interaction reward value is written into the attitude update sample record as a feature of the chassis balance state. The environmental interaction reward value is used to characterize the degree to which the all-terrain adaptive mobile chassis approaches the reference horizontal state after performing the above-mentioned telescopic action. The attitude update sample record also includes terrain elevation feedback sequence, terrain change feature, first hidden state representation, second hidden state representation, spliced ​​features, target vehicle attitude angle, swing arm adjustment angle, telescopic adjustment length, and current control cycle marker, so that the environmental interaction reward value can establish a homologous correspondence with the input and output data of the preceding attitude decision network. When calculating the time difference error, first, the current cumulative environmental interaction reward prediction value generated by the attitude decision network based on the stitched features to generate the target vehicle attitude angle under the current control cycle mark is read. Then, the environmental interaction reward value generated after performing the above-mentioned scaling action is read, and combined with the terrain elevation feedback sequence and terrain abrupt change features of the next control cycle, the next cumulative environmental interaction reward prediction value is formed. Subsequently, the environmental interaction reward value and the next cumulative environmental interaction reward prediction value are subjected to attenuation correlation processing to form an updated reference reward value. The difference between the updated reference reward value and the current cumulative environmental interaction reward prediction value is calculated to form the time difference error. The current cumulative environmental interaction reward prediction value is used to represent the prediction result of the attitude decision network for the environmental interaction reward value in the subsequent multiple control cycles under the current control cycle mark. The next cumulative environmental interaction reward prediction value is used to represent the prediction result of the attitude decision network for the environmental interaction reward value in the subsequent multiple control cycles in the next control cycle. The updated reference reward value is used to include the environmental interaction reward value that has already occurred and the prediction result of the next control cycle in the same time difference error calculation caliber. The time difference error is used to express the degree of deviation of the attitude decision network from the prediction of the subsequent chassis balance state under the current control cycle mark. It is not a simple angle error, but a difference after incorporating the target vehicle attitude angle, swing arm adjustment angle, telescopic adjustment length, and environmental interaction reward value after performing the above telescopic actions into the same update caliber.When calculating gradient information based on temporal difference error, the temporal difference error is first propagated backward along the parameter propagation path of the policy output layer, the concatenated features, the state-aware encoding layer, and the spatial attention layer. Then, the direction of change of the temporal difference error relative to the temporal feature encoding parameters in the state-aware encoding layer, the environmental obstacle weight allocation parameters in the spatial attention layer, and the attitude angle candidate encoding parameters in the policy output layer are calculated to form gradient information. This gradient information indicates the direction and magnitude of adjustment that the network mapping parameters should be in different structural layers, providing an executable numerical basis for subsequent iterative updates of the network mapping parameters.

[0058] Preferably, the specific implementation process of the step "iteratively updating the network mapping parameters in the attitude decision network according to the descent direction corresponding to the gradient information to maximize the cumulative environmental interaction reward value generated by subsequent actions" is as follows: After the gradient information has been formed, the network mapping parameters in the attitude decision network are first read, and the network mapping parameters are divided into state-aware encoding layer mapping parameters, spatial attention layer mapping parameters, and policy output layer mapping parameters according to the structural layer where the network mapping parameters are located. The state-aware encoding layer mapping parameters are used to control the mapping relationship of the terrain elevation feedback sequence after temporal feature encoding to form the first hidden state representation. The spatial attention layer mapping parameters are used to control the mapping relationship of the terrain abrupt feature after environmental obstacle weight allocation to form the second hidden state representation. The policy output layer mapping parameters are used to control the mapping relationship of the spliced ​​features after attitude angle candidate encoding and attitude constraint screening to form the target vehicle attitude angle. The above network mapping parameters jointly determine how the terrain elevation feedback sequence and terrain abrupt feature affect the target vehicle attitude angle. When iteratively updating the network mapping parameters according to the descent direction corresponding to the gradient information, the gradient information is first matched to the mapping parameters of the state-aware encoding layer, the spatial attention layer, and the policy output layer, respectively. Then, the mapping parameters of the state-aware encoding layer, the spatial attention layer, and the policy output layer are corrected layer by layer according to the descent direction corresponding to the temporal difference error to form the updated network mapping parameters. The descent direction is used to make the temporal difference error decrease in subsequent control cycles, that is, to make the target vehicle attitude angle generated by the attitude decision network more closely match the environmental interaction reward value formed after performing the above-mentioned scaling action. After completing the layer-by-layer correction, the updated network mapping parameters are written into the network parameter update record. The network parameter update record includes the current control cycle marker, the temporal difference error, the gradient information, the network mapping parameters before the update, and the updated network mapping parameters. Subsequently, the attitude decision network reads the updated network mapping parameters in the next control cycle and redetermines the target vehicle attitude angle based on the terrain elevation feedback sequence and terrain change features of the next control cycle. The cumulative environmental interaction reward value is used to represent the overall attitude response evaluation result formed by the temporal correlation of environmental interaction reward values ​​within multiple consecutive control cycles. It reflects the trend of chassis balance state change after the all-terrain adaptive mobile chassis performs swing arm action and telescopic action multiple times in continuously abrupt terrain.By iteratively updating the network mapping parameters according to the descent direction corresponding to the gradient information, the state-aware encoding layer can gradually adjust the encoding method of the height change trend in the terrain elevation feedback sequence, the spatial attention layer can gradually adjust the weight allocation method of the local abrupt change region in the terrain abrupt change feature, and the policy output layer can gradually adjust the mapping method of generating the target vehicle body attitude angle from the spliced ​​features. As a result, the cumulative environmental interaction reward value generated by subsequent actions tends to increase in the continuous control process, and the attitude adjustment process of the all-terrain adaptive mobile chassis under abrupt terrain changes is more in line with the chassis balance state feedback after execution.

[0059] Optionally, the step of constructing the environmental interaction reward value based on the first deviation value of the current pitch angle of the chassis from the horizontal plane and the second deviation value of the current roll angle of the chassis from the horizontal plane includes: The first deviation value, which represents the current pitch angle of the chassis deviating from the horizontal plane, and the second deviation value, which represents the current roll angle of the chassis deviating from the horizontal plane, are respectively reverse-mapped to obtain a first basic reward item and a second basic reward item that are inversely proportional to the degree of deviation. The peak value of the transient torque change fed back by each of the wheel-leg composite units during the above-mentioned extension and retraction action is obtained, and then a penalty term for suppressing the intensity of the action is constructed based on the peak value of the transient torque change. The first basic reward item and the second basic reward item are fused to obtain the basic reward index; By introducing the penalty term to reduce the weight of the basic reward indicator, the environmental interaction reward value is constructed based on the penalty term after the weight reduction.

[0060] Preferably, the specific implementation process of the step "reverse mapping the first deviation value representing the current pitch angle of the chassis deviating from the horizontal plane and the second deviation value representing the current roll angle of the chassis deviating from the horizontal plane to obtain the first basic reward item and the second basic reward item that are inversely proportional to the degree of deviation" is as follows: After the current pitch angle and the current roll angle of the chassis have been written into the chassis attitude feedback record, the current pitch angle, the current roll angle, the current control cycle marker, and the extension / retraction action completion status of the chassis attitude feedback record are read first, and the reference horizontal state corresponding to the all-terrain adaptive mobile chassis is read. The reference horizontal state is configured when the all-terrain adaptive mobile chassis completes the assembly calibration or starts the initialization calibration. During configuration, the pitch reference attitude and roll reference attitude when the vehicle body reference plane is stably supported on the horizontal support surface are used as the attitude comparison reference. The attitude comparison reference is established with the installation direction of the inertial measurement unit, the longitudinal coordinate direction and the lateral coordinate direction in the preset chassis global coordinate system, so that the subsequently read current pitch angle and current roll angle of the chassis can participate in the deviation calculation under the same attitude comparison reference. Subsequently, the difference between the current pitch angle of the chassis and the pitch reference attitude in the reference horizontal state is calculated using the same dimension to form a first deviation value; the difference between the current roll angle of the chassis and the roll reference attitude in the reference horizontal state is calculated using the same dimension to form a second deviation value. The first deviation value is used to represent the degree of forward and backward tilt deviation of the vehicle reference plane in the longitudinal direction after performing the above-mentioned telescopic action, and the second deviation value is used to represent the degree of left and right tilt deviation of the vehicle reference plane in the lateral direction after performing the above-mentioned telescopic action; the first deviation value and the second deviation value together reflect the source of attitude deviation directly related to vehicle attitude offset in the chassis balance state characteristics. When performing reverse mapping, the pre-configured reward mapping configuration record is first read. The reward mapping configuration record is configured before the operation of the all-terrain adaptive mobile chassis according to the allowable pitch deviation range, the allowable roll deviation range, the measurement resolution capability of the inertial measurement unit, and the attitude adjustment range of the wheel-leg composite unit; the reward mapping configuration record is used to convert the first deviation value and the second deviation value with angular dimensions into dimensionless reward evaluation data, avoiding the direct entry of angular deviation into the environmental interaction reward value and causing dimensional mixing between different attitude axes. Subsequently, based on the reward mapping configuration record, the first deviation value and the second deviation value are normalized to form first deviation normalized data and second deviation normalized data. The first deviation normalized data is used to represent the relative deviation of the first deviation value within the allowable pitch deviation range, and the second deviation normalized data is used to represent the relative deviation of the second deviation value within the allowable roll deviation range. Then, according to the mapping relationship that the smaller the deviation, the higher the reward contribution, the first deviation normalized data is reverse-mapped to obtain the first basic reward item, and the second deviation normalized data is reverse-mapped to obtain the second basic reward item.The first basic reward item expresses the degree to which the all-terrain adaptive mobile chassis approaches the reference horizontal state in the longitudinal direction of travel, and the second basic reward item expresses the degree to which the all-terrain adaptive mobile chassis approaches the reference horizontal state in the lateral direction. Both the first and second basic reward items belong to the reward evaluation data. The first and second basic reward items subsequently enter the feature fusion process of the basic reward index together, so that the environmental interaction reward value can simultaneously reflect the pitch attitude response and roll attitude response after performing the above-mentioned telescopic action, rather than evaluating the adjustment result of the wheel-leg composite unit based on only a single attitude axis. The inverse mapping relationship inversely proportional to the degree of deviation is adopted because the closer the current pitch angle and current roll angle of the chassis are to the reference horizontal state, the closer the attitude adjustment result formed by the previous target vehicle attitude angle, swing arm adjustment angle, and telescopic adjustment length in the current control cycle is to the terrain passage requirements. Conversely, when the first deviation value or the second deviation value is large, the first or second basic reward item is reduced accordingly, thereby providing numerical feedback directly related to the chassis balance state characteristics for the subsequent iterative update of network mapping parameters in the attitude decision network.

[0061] Preferably, the specific implementation process of the step "obtaining the peak value of the transient torque change fed back by each of the wheel-leg composite units during the execution of the above-mentioned telescopic action, and then constructing a penalty term for suppressing the severity of the action based on the peak value of the transient torque change" is as follows: After the first basic reward term and the second basic reward term have been formed, the telescopic action issuance record, the swing arm action issuance record, and the execution status feedback record formed by the corresponding wheel-leg composite unit under the current control cycle mark are read. The execution status feedback record includes not only the current rotary joint angle, the current linear actuator length, and the wheel end grounding status feedback, but also the torque feedback sequence of the wheel-leg composite unit during the execution of the above-mentioned telescopic action; the torque feedback sequence comes from the drive load feedback of the rotary joint actuator and the linear drive actuator during the action execution process, and the torque feedback sequence is used to reflect the changes in ground reaction force, structural force, and drive load experienced by the wheel-leg composite unit during the connection between the swing arm action and the telescopic action. The reason for obtaining the torque feedback sequence during the above-mentioned telescopic action is that the current pitch angle and roll angle of the chassis can only reflect the attitude result after the telescopic action is completed, and cannot reflect whether there has been a large transient force change during the execution process. In abrupt terrain such as gravel, step edges, and ditch edges, even if the final current pitch angle and roll angle of the chassis are close to the reference level, if a high peak value of transient torque change occurs during the above-mentioned telescopic action, it also indicates that the wheel-leg composite unit has experienced a violent support switch or ground impact. Such execution processes also need to be included in the construction caliber of environmental interaction reward value. When generating the torque transient peak, the torque feedback sequence is first time-aligned according to the current control cycle marker, so that the torque feedback sequence corresponds to the first drive command stream, the second drive command stream, and the completion status of the telescopic action to the same control cycle. Then, the torque feedback sequence undergoes sampling point sequence verification and abnormal jump point removal to form an effective torque feedback sequence. This effective torque feedback sequence is used to exclude non-action load changes caused by communication jitter or sampling loss. Next, the rate of change of torque between adjacent sampling points in the effective torque feedback sequence is extracted to form torque rate of change data. The maximum abrupt change amplitude related to the aforementioned telescopic action is selected from the torque rate of change data to form the torque transient peak. The torque transient peak is not a single static torque value, but rather a peak value representing the degree of load change drastic by the wheel-leg composite unit during the aforementioned telescopic action. The higher the torque transient peak, the greater the impact or support change caused by the swing arm action and telescopic action of the wheel-leg composite unit in the current control cycle.When constructing the penalty term, the pre-configured torque evaluation configuration record is first read. This record is configured during assembly calibration or initialization calibration based on the allowable torque range of the rotary joint actuator, the allowable torque range of the linear drive actuator, the motion direction constraints of the wheel-leg composite unit, and the tolerance range of the vehicle body reference plane to local impacts. The torque evaluation configuration record includes a torque reference range, which is used to convert the peak value of transient torque changes into motion process evaluation data under the same evaluation caliber as the basic reward index. Subsequently, the peak value of transient torque changes and the torque reference range in the torque evaluation configuration record are normalized to form torque change evaluation data. A penalty term is then constructed based on this torque change evaluation data. This penalty term is used to numerically suppress the severity of the motion. It subsequently participates in the weighting process of the basic reward index, ensuring that the environmental interaction reward value evaluates both the chassis attitude result after performing the aforementioned telescopic motion and the smoothness of the motion during the telescopic motion process. This avoids the attitude decision network only pursuing a reduction in attitude angle deviation while ignoring load changes during the execution of the wheel-leg composite unit.

[0062] Preferably, in the specific technical implementation of the step "performing feature fusion on the first basic reward item and the second basic reward item to obtain basic reward indicators", after the first basic reward item, the second basic reward item, and the penalty item have all been formed, the first deviation value corresponding to the first basic reward item, the second deviation value corresponding to the second basic reward item, the current control cycle marker, and the structural installation topology relationship of the all-terrain adaptive mobile chassis are read first. Then, the attitude evaluation weights of the longitudinal travel direction and the lateral direction in the current control cycle are determined based on the structural installation topology relationship. The attitude evaluation weights can be pre-configured in the reward mapping configuration record, or they can be adjusted in association with the contour geometry and spatial distribution density of abrupt change points in the current terrain abrupt change features. For example, when the terrain abrupt change features show that the local height abrupt change area mainly extends along the lateral direction, the numerical contribution of the second basic reward item in feature fusion can be increased accordingly; when the terrain abrupt change features show that the local height abrupt change area is mainly located in front of the travel direction, the numerical contribution of the first basic reward item in feature fusion can be increased accordingly. The aforementioned attitude evaluation weights do not perform semantic judgment on the all-terrain adaptive mobile chassis operation scenario. Instead, based on the spatial distribution relationship of terrain change features in a preset global chassis coordinate system, they perform numerical matching processing on the first and second basic reward items along the attitude axis dimension, ensuring that the basic reward indicators correspond to the direction of the impact of the current terrain change on the vehicle's attitude. During feature fusion, the first basic reward item is first written into the longitudinal attitude reward field, and the second basic reward item into the lateral attitude reward field. The longitudinal attitude reward field carries the numerical contribution of the first basic reward item to the pitch attitude response, and the lateral attitude reward field carries the numerical contribution of the second basic reward item to the roll attitude response. Then, the longitudinal and lateral attitude reward fields are weighted and fused according to the attitude evaluation weights to form comprehensive attitude reward data. This comprehensive attitude reward data expresses the degree to which the all-terrain adaptive mobile chassis approaches a baseline horizontal state in both the pitch and roll directions after performing the aforementioned telescopic action. It does not simply select the larger value of the first or second basic reward item, but rather retains the influence of both longitudinal and lateral attitude deviations under the same current control cycle marker. Subsequently, the overall attitude reward data undergoes deviation balancing correction processing to form a basic reward index. This deviation balancing correction processing is used to prevent the basic reward index from remaining too high when the first deviation value is small and the second deviation value is large, and also to prevent the basic reward index from remaining too high when the second deviation value is small and the first deviation value is large. In other words, the basic reward index needs to reflect the overall state of the vehicle body reference plane approaching the reference horizontal state, rather than just reflecting the local good state of a certain attitude axis.The reason for fusing the first and second basic reward items is that the attitude stability of the all-terrain adaptive mobile chassis under abrupt terrain changes is simultaneously affected by pitch changes in the longitudinal direction of travel and roll changes in the lateral direction. The preceding target vehicle attitude angle also usually includes pitch adjustment direction and roll adjustment direction. If the environmental interaction reward value is constructed only based on the first or second basic reward item, the iterative update of the network mapping parameters in the attitude decision network will be biased towards a single attitude axis, thereby reducing the adaptability of the subsequent target vehicle attitude angle to complex terrain spatial undulations. By forming a basic reward index, the first and second basic reward items are incorporated into the same attitude evaluation caliber. The basic reward index is then subjected to a weight reduction process by a penalty item, enabling the environmental interaction reward value to form a continuous technical connection between attitude result evaluation and motion process evaluation.

[0063] Preferably, the specific implementation process of the step "reducing the weight of the basic reward index by introducing the penalty term, and then constructing the environmental interaction reward value" is as follows: After the basic reward index has been formed by feature fusion of the first and second basic reward terms, and the penalty term has been formed by constructing the torque transient change peak, the basic reward index, penalty term, current control cycle marker, and extension / retraction action completion status are read first, and the above data are written into the reward construction record. The reward construction record is used to include the attitude result evaluation after performing the above extension / retraction action and the action intensity evaluation during the above extension / retraction action into the same construction caliber, so that the environmental interaction reward value can correspond to the same target vehicle attitude angle generation, swing arm adjustment angle execution, and extension / retraction length execution process within the current control cycle. When performing weight reduction processing, the reward weight reduction magnitude is first determined based on the penalty term. The reward weight reduction magnitude is used to represent the degree of suppression of the torque transient change peak on the basic reward index; then, the reward weight reduction magnitude is applied to the basic reward index to form the weighted basic reward index. The reduced-weighted basic reward index is not the penalty item itself, but rather the reward evaluation result obtained after introducing the penalty item. This processing relationship avoids using the intensity of the action as the environmental interaction reward value alone, and also avoids evaluating the attitude adjustment result solely using the current chassis pitch angle and current chassis roll angle. Specifically, when constructing the environmental interaction reward value, the reduced-weighted basic reward index is first subject to boundary limiting processing to keep it within the reward output range allowed by the reward mapping configuration record. This boundary limiting processing constrains the reduced-weighted basic reward index within the reward output range, which is derived from the reward mapping configuration record and matched with the subsequent time difference error calculation caliber of the environmental interaction reward value. Then, the reduced-weighted basic reward index after boundary limiting processing is associated with the current control cycle marker, target vehicle attitude angle, swing arm adjustment angle, telescopic adjustment length, and chassis balance state characteristics to generate the environmental interaction reward value. The environmental interaction reward value represents the comprehensive attitude feedback evaluation formed by the all-terrain adaptive mobile chassis after performing swing arm and telescopic movements in the current control cycle. It simultaneously includes the longitudinal attitude evaluation corresponding to the first deviation value, the lateral attitude evaluation corresponding to the second deviation value, and the motion process suppression evaluation corresponding to the peak value of torque transient change. The reason for reducing the weight of the basic reward index through a penalty term is that under abrupt terrain changes, the generated target vehicle attitude angle must not only make the vehicle reference plane approach the reference horizontal state, but also prevent the wheel-leg composite unit from generating excessively high transient load changes when performing swing arm and telescopic movements. If the environmental interaction reward value is constructed based solely on the basic reward index, the network mapping parameters of the attitude decision network are prone to bias towards rapidly reducing attitude deviations during iterative updates, while ignoring the motion impact reflected by the peak value of torque transient change.After reducing the weight of the basic reward index by the penalty term, the environmental interaction reward value can feed back the improvement of posture deviation and the suppression of the intensity of movement to the time difference error calculation process, and further participate in the formation of gradient information. This enables the posture decision network to simultaneously absorb the actual execution results expressed by the chassis balance state characteristics and the changes in the motion load of the wheel-leg composite unit when adjusting the network mapping parameters in subsequent control cycles.

[0064] Optionally, the terrain elevation feedback sequence includes local detection distance data acquired by the multiple wheel-leg composite units respectively through their built-in non-contact ranging probes; acquiring the terrain elevation feedback sequence of the multiple independently driven wheel-leg composite units on the all-terrain adaptive mobile chassis at the current moment includes: At the current moment, the non-contact ranging probes built into the multiple independently driven wheel-leg composite units on the all-terrain adaptive mobile chassis are synchronously triggered to perform local ground ranging. Acquire the local detection distance data returned by each of the non-contact ranging probes; The local detection distance data are sorted and combined according to the structural installation topology of the wheel-leg composite unit on the chassis to form a terrain elevation feedback sequence covering the grounding area around the all-terrain adaptive mobile chassis.

[0065] Preferably, the specific implementation process of step "synchronously triggering the non-contact ranging probes of the multiple independently driven wheel-leg composite units on the all-terrain adaptive mobile chassis to perform local ground ranging" is as follows: When the all-terrain adaptive mobile chassis enters a chassis control cycle, the sampling trigger configuration record is read first. The sampling trigger configuration record is configured when the all-terrain adaptive mobile chassis completes assembly calibration or starts initialization calibration. The sampling trigger configuration record includes the wheel-leg composite unit installation node corresponding to each of the multiple independently driven wheel-leg composite units, the installation offset of the non-contact ranging probe, the ranging direction of the non-contact ranging probe, the trigger beat of the chassis control cycle, the data arrangement caliber of the terrain elevation feedback sequence, and the effective sampling time window of the local detection distance data. The current time is not an arbitrary reading point, but a sampling trigger time jointly determined by the trigger beat of the chassis control cycle, the current travel state of the all-terrain adaptive mobile chassis, and the completion status of the extension and retraction actions of the previous chassis control cycle. Before the swing arm action and extension and retraction actions of the previous chassis control cycle are completed and before the all-terrain adaptive mobile chassis enters the next attitude judgment, a current control cycle marker is generated according to the sampling trigger configuration record and written into the sampling trigger record so that the local detection distance data returned by the subsequent non-contact ranging probes can correspond to the same sampling trigger time. Subsequently, based on the current control cycle marker in the sampling trigger record and the effective sampling time window of the local detection distance data, ranging trigger commands are simultaneously issued to the non-contact ranging probes built into multiple independently driven wheel-leg composite units, so that each non-contact ranging probe begins local ground ranging within the same chassis control cycle. The reason for using synchronous triggering is that when the all-terrain adaptive mobile chassis travels in complex environments such as steps, gravel, slopes, snow, and soft ground, the ground contact areas corresponding to different wheel-leg composite units will continuously change with the vehicle's attitude. If multiple non-contact ranging probes measure distances separately at different sampling trigger times, the local detection distance data corresponding to the front ground contact area and the local detection distance data corresponding to the rear ground contact area will correspond to different vehicle attitude states. The subsequent terrain elevation feedback sequence is prone to misinterpreting time changes as spatial height differences. By unifying the current control cycle marker through sampling trigger records and synchronously triggering multiple non-contact ranging probes to perform local ground ranging through ranging trigger commands, the local detection distance data of the ground contact areas corresponding to multiple wheel-leg composite units can have the same temporal caliber, thus providing a temporally consistent input basis for subsequent spatial distribution analysis and terrain change feature extraction.

[0066] Preferably, in the specific technical implementation of the step "acquiring the local detection distance data returned by each of the non-contact ranging probes", after multiple non-contact ranging probes have completed local ground ranging according to the ranging trigger command, the probe return data record returned by each non-contact ranging probe is first read according to the current control cycle mark in the sampling trigger record. The probe return data record includes the wheel-leg composite unit mounting node, the non-contact ranging probe mounting offset, the non-contact ranging probe ranging direction, the ranging echo timestamp, the ranging echo intensity, the original ranging reading, and the current control cycle mark; the probe return data record is used to establish a correspondence between the original ranging reading and the wheel-leg composite unit mounting node that generated the original ranging reading, the non-contact ranging probe ranging direction, and the current control cycle mark, so that the subsequently read local detection distance data can be traced back to the corresponding non-contact ranging probe and the corresponding grounding area. When acquiring local detection distance data, the probe's returned data records are first matched with the current control cycle marker, and probe returned data records that do not belong to the current control cycle marker are removed to form the current control cycle distance measurement record. Subsequently, the distance echo intensity and the original distance reading in the current control cycle distance measurement record are validated to eliminate invalid distance readings caused by insufficient reflection from soft ground, abrupt echo changes at the edge of gravel, or scattering from the snow surface, thus forming a valid distance measurement record. The valid distance measurement record continues to retain the wheel-leg composite unit installation node, the non-contact distance measuring probe installation offset, the non-contact distance measuring probe distance direction, and the current control cycle marker, and converts the original distance reading that has passed the validity validation into local detection distance data. The local detection distance data does not directly represent the coordinate data of terrain elevation, but rather the distance information between the non-contact ranging probe along its measuring direction and the local ground reflection point. This local detection distance data needs to be combined with the non-contact ranging probe's installation offset, measuring direction, and the wheel-leg composite unit's mounting node to be converted into height feedback corresponding to the ground contact area around the all-terrain adaptive mobile chassis. This is because the non-contact ranging probe is mounted on the wheel-leg composite unit, and the local detection distance data returned by it is affected by the current mounting position of the wheel-leg composite unit, the measuring direction of the non-contact ranging probe, and the vehicle's reference plane attitude. By preserving this correspondence between the probe's returned data records and the valid ranging records, the system avoids using individual raw ranging readings independently of the wheel-leg composite unit's mounting node, thus enabling the local detection distance data to serve as a reliable input field for the terrain elevation feedback sequence in subsequent sorting and combination.

[0067] Preferably, the specific implementation process of the step "sorting and combining the local detection distance data according to the structural installation topology of the wheel-leg composite unit on the chassis" is as follows: After the local detection distance data returned by each non-contact ranging probe has been formed, the structural installation topology configuration record is read first. The structural installation topology configuration record is configured when the all-terrain adaptive mobile chassis completes assembly calibration or starts initialization calibration. The structural installation topology configuration record includes the wheel-leg composite unit installation nodes corresponding to each of the multiple independently driven wheel-leg composite units, the installation offset of the wheel-leg composite unit installation nodes in the preset chassis global coordinate system, the front-rear position relationship of the wheel-leg composite unit installation nodes relative to the chassis geometric center, the left-right position relationship of the wheel-leg composite unit installation nodes relative to the chassis geometric center, the installation offset of the non-contact ranging probe, and the data arrangement caliber of the terrain elevation feedback sequence. The structural installation topology is not simply a hardware numbering sequence, but rather represents the spatial installation relationship of multiple independently driven wheel-leg composite units on an all-terrain adaptive mobile chassis. This topology maps the local detection distance data corresponding to different wheel-leg composite units to the front, rear, left, right, and corner grounding areas of the all-terrain adaptive mobile chassis, ensuring that subsequent terrain elevation feedback sequences reflect the spatial distribution of the grounding areas around the chassis. During sorting and combination, the wheel-leg composite unit installation nodes in the configuration record are first configured according to the structural installation topology. Then, node matching processing is performed on the local detection distance data in the effective ranging records to form node-matched local detection distance data. This node-matched local detection distance data retains the current control cycle marker, wheel-leg composite unit installation node, non-contact ranging probe installation offset, and non-contact ranging probe ranging direction, ensuring that each node-matched local detection distance data corresponds to a wheel-leg composite unit installation node. Subsequently, based on the data arrangement caliber of the terrain elevation feedback sequence, the local detection distance data after node matching is sorted according to the front-to-back and left-to-right positional relationships of the wheel-leg composite unit installation nodes in the preset chassis global coordinate system to form topology-sorted distance data. This topology-sorted distance data is used to organize the originally scattered local detection distance data into a data representation with a fixed spatial order, enabling subsequent spatial distribution analysis to identify elevation changes between adjacent grounding areas in the same order.The reason for sorting and combining the data by structural installation topology is that when local detection distance data of the same value is located in the front grounding area, rear grounding area, left grounding area, or right grounding area, the direction of influence on the target vehicle attitude angle is not the same. Only by first unifying the spatial order of local detection distance data according to structural installation topology can the subsequent continuous terrain point cloud contour, elevation change gradient, and terrain change features be consistent with the actual grounding area of ​​the all-terrain adaptive mobile chassis.

[0068] Preferably, the specific implementation process of step "to form the terrain elevation feedback sequence covering the grounding area around the all-terrain adaptive mobile chassis" is as follows: After the topology sorting distance data is formed, the local detection distance data in the topology sorting distance data is first converted to a higher caliber based on the installation offset and ranging direction of the non-contact ranging probe in the structural installation topology configuration record, to form node elevation feedback data. The node elevation feedback data is used to represent the local height feedback of the grounding area corresponding to each wheel leg composite unit installation node relative to the vehicle body reference plane. The node elevation feedback data retains the ranging source of the local detection distance data and converts the local detection distance data into a height field that can participate in the construction of the terrain elevation feedback sequence. Subsequently, according to the data arrangement caliber of the terrain elevation feedback sequence in the structural installation topology configuration record, multiple node elevation feedback data are serialized and encapsulated according to the topology sorting result of the wheel leg composite unit installation nodes to form a terrain elevation feedback sequence. The terrain elevation feedback sequence includes the current control cycle marker, node elevation feedback data corresponding to each of the multiple wheel-leg composite unit installation nodes, the spatial arrangement order of the node elevation feedback data, and the effective distance measurement marker of the node elevation feedback data. The terrain elevation feedback sequence is used to express the height distribution of the ground contact area around the all-terrain adaptive mobile chassis at the current moment, rather than the independent distance measurement results near a single wheel-leg composite unit. To ensure that the terrain elevation feedback sequence covers the ground contact area around the all-terrain adaptive mobile chassis, spatial adjacency markers can be applied to the node elevation feedback data between adjacent wheel-leg composite unit installation nodes based on the installation offset of the wheel-leg composite unit installation nodes in a preset global chassis coordinate system, thus forming a ground contact area adjacency relationship. This ground contact area adjacency relationship is written into the terrain elevation feedback sequence generation record along with the terrain elevation feedback sequence. The terrain elevation feedback sequence generation record is used to include the current control cycle marker, node elevation feedback data, spatial arrangement order of the node elevation feedback data, effective distance measurement marker of the node elevation feedback data, and ground contact area adjacency relationship into the same data source caliber. The resulting terrain elevation feedback sequence can directly participate in the generation of continuous terrain point cloud contours in subsequent spatial distribution analysis, and continue to participate in the extraction of elevation change gradients, the marking of abrupt change candidate points, and the generation of terrain abrupt change features. By converting local detection distance data into node elevation feedback data, and then serializing and encapsulating the node elevation feedback data according to the structural installation topology, the terrain elevation feedback sequence transforms the local ranging results of the non-contact ranging probe into the spatial height input required for all-terrain adaptive mobile chassis attitude control. This ensures that the subsequent formation of the target vehicle attitude angle, swing arm adjustment angle, and telescopic adjustment length can all be based on the actual height feedback of the surrounding ground area at the current moment.

[0069] like Figure 3As shown, this is a control device for an all-terrain adaptive mobile chassis according to an embodiment of this application, which includes: The feedback acquisition module is used to acquire the terrain elevation feedback sequence of multiple independently driven wheel-leg composite units on the all-terrain adaptive mobile chassis at the current moment. The feature extraction module is used to perform spatial distribution analysis on the terrain elevation feedback sequence in order to extract the terrain abrupt change features of the current environment; The attitude determination module is used to input the terrain elevation feedback sequence and the terrain abrupt change features into a pre-configured attitude decision network to determine the target vehicle attitude angle of the all-terrain adaptive mobile chassis. The adjustment parameter calculation module is used to calculate the swing arm adjustment angle and telescopic adjustment length of each of the multiple independently driven wheel-leg composite units according to the target vehicle body attitude angle. The motion control module is used to control the corresponding wheel-leg composite unit to perform the swing arm motion according to the swing arm adjustment angle, and after the swing arm motion is completed, control the corresponding wheel-leg composite unit to perform the telescopic motion according to the telescopic adjustment length. The parameter update module is used to iteratively update the network mapping parameters in the attitude decision network based on the chassis balance state characteristics after performing the above actions.

[0070] like Figure 4 As shown, an electronic device according to an embodiment of this application includes: At least one processor; and, A memory that is communicatively connected to the at least one processor; The memory stores a computer program that can be executed by the at least one processor. When the computer program is executed by the at least one processor, it causes the at least one processor to perform the control method for the all-terrain adaptive mobile chassis as described in this application.

[0071] like Figure 5 As shown, this is a computer-readable storage medium according to an embodiment of the present application, wherein the computer-readable storage medium stores a computer program; When the computer program is executed by the processor, it implements the control method for the all-terrain adaptive mobile chassis as described in this application.

[0072] like Figure 6 As shown, a mobile robot according to an embodiment of this application includes: The robot itself; An all-terrain adaptive mobile chassis is mounted on the robot body, and the all-terrain adaptive mobile chassis includes multiple independently driven wheel-leg composite units; The terrain elevation feedback acquisition component is used to acquire the terrain elevation feedback sequence of the multiple independently driven wheel-leg composite units at the current moment; The chassis controller is communicatively connected to the terrain elevation feedback acquisition component and the multiple independently driven wheel-leg composite units; The chassis controller is used to perform spatial distribution analysis on the terrain elevation feedback sequence in order to extract the terrain abrupt change features of the current environment; The chassis controller is also used to input the terrain elevation feedback sequence and the terrain abrupt change features into a pre-configured attitude decision network to determine the target vehicle attitude angle of the all-terrain adaptive mobile chassis. The chassis controller is also used to calculate the swing arm adjustment angle and telescopic adjustment length of each of the multiple independently driven wheel-leg composite units according to the target vehicle body attitude angle. The chassis controller is also used to control the corresponding wheel-leg composite unit to perform the swing arm action according to the swing arm adjustment angle, and after the swing arm action is completed, control the corresponding wheel-leg composite unit to perform the telescopic action according to the telescopic adjustment length. The chassis controller is also used to iteratively update the network mapping parameters in the attitude decision network based on the chassis balance state characteristics after performing the above actions.

[0073] like Figure 7 As shown, this is a control system for an all-terrain adaptive mobile chassis according to an embodiment of this application, which includes: An all-terrain adaptive mobile chassis, comprising multiple independently driven wheel-leg composite units; A terrain elevation feedback acquisition device is used to acquire the terrain elevation feedback sequence of the multiple independently driven wheel-leg composite units at the current moment; The attitude control device is communicatively connected to the terrain elevation feedback acquisition device and the multiple independently driven wheel-leg composite units; The attitude control device is used to perform spatial distribution analysis on the terrain elevation feedback sequence in order to extract the terrain abrupt change features of the current environment. The attitude control device is also used to input the terrain elevation feedback sequence and the terrain abrupt change features into a pre-configured attitude decision network to determine the target vehicle attitude angle of the all-terrain adaptive mobile chassis. The attitude control device is also used to calculate the swing arm adjustment angle and telescopic adjustment length of each of the multiple independently driven wheel-leg composite units according to the target vehicle body attitude angle. The attitude control device is also used to control the corresponding wheel-leg composite unit to perform the swing arm action according to the swing arm adjustment angle, and after the swing arm action is completed, control the corresponding wheel-leg composite unit to perform the telescopic action according to the telescopic adjustment length. The attitude control device is also used to iteratively update the network mapping parameters in the attitude decision network based on the chassis balance state characteristics after performing the above actions.

[0074] Figures 2-7 For an exemplary description, please refer to the above. Figure 1 This will not be elaborated upon here.

Claims

1. A control method for an all-terrain adaptive mobile chassis, characterized in that, include: Obtain the terrain elevation feedback sequence of multiple independently driven wheel-leg composite units on the all-terrain adaptive mobile chassis at the current moment; Spatial distribution analysis is performed on the terrain elevation feedback sequence to extract the terrain abrupt change features of the current environment; The terrain elevation feedback sequence and the terrain abrupt change features are input into a pre-configured attitude decision network to determine the target vehicle attitude angle of the all-terrain adaptive mobile chassis. Based on the target vehicle body attitude angle, calculate the swing arm adjustment angle and telescopic adjustment length of each of the multiple independently driven wheel-leg composite units; The corresponding wheel-leg composite unit is controlled to perform the swing arm action according to the swing arm adjustment angle, and after the swing arm action is completed, the corresponding wheel-leg composite unit is controlled to perform the telescopic action according to the telescopic adjustment length. The network mapping parameters in the attitude decision network are iteratively updated based on the chassis balance state characteristics after performing the above-mentioned telescopic action.

2. The control method for an all-terrain adaptive mobile chassis according to claim 1, characterized in that, The step of spatially distributing the terrain elevation feedback sequence to extract the terrain abrupt change features of the current environment includes: The terrain elevation feedback sequence is mapped to a preset chassis global coordinate system to generate a continuous terrain point cloud outline that represents the terrain undulation state at the current moment. The spatial distribution is analyzed in the continuous terrain point cloud outline to extract the elevation change gradient between adjacent coordinate points; The coordinate points where the elevation change gradient exceeds a preset gradient threshold are marked as abrupt change candidate points; The multiple mutation candidate points are subjected to regional feature aggregation processing to generate the terrain mutation feature containing the spatial distribution density and contour geometry of the mutation candidate points.

3. The control method for an all-terrain adaptive mobile chassis according to claim 1, characterized in that, The step of inputting the terrain elevation feedback sequence and the terrain abrupt change features into a pre-configured attitude decision network to determine the target vehicle attitude angle of the all-terrain adaptive mobile chassis includes: The terrain elevation feedback sequence is temporally featured by the state-aware coding layer in the attitude decision network to obtain the first hidden state representation. The spatial attention layer in the posture decision network is used to assign environmental obstacle weights to the terrain abruptness features to obtain the second hidden state representation. The first hidden state representation and the second hidden state representation are concatenated, and then the concatenated features are input into the policy output layer of the attitude decision network to generate the target vehicle attitude angle of the all-terrain adaptive mobile chassis to maintain a baseline level state under the corresponding terrain.

4. The control method for an all-terrain adaptive mobile chassis according to claim 1, characterized in that, The step of calculating the swing arm adjustment angle and telescopic adjustment length corresponding to each of the multiple independently driven wheel-leg composite units based on the target vehicle body attitude angle includes: Obtain the initial physical configuration parameters of the multiple independently driven wheel-leg composite units at the current moment; Construct the inverse kinematics transformation matrix of the all-terrain adaptive mobile chassis based on the target vehicle body attitude angle; Substitute the initial physical configuration parameters into the inverse kinematics transformation matrix to obtain the expected relative coordinate positions of each wheel-leg composite unit relative to the all-terrain adaptive mobile chassis. Based on the desired relative coordinate position, the expected rotation value of the rotary joint and the expected stroke value of the linear actuator corresponding to each of the wheel-leg composite units to reach the coordinate position are calculated respectively. Then, the expected rotation value of the rotary joint is used as the adjustment angle of the swing arm, and the expected stroke value of the linear actuator is used as the extension adjustment length.

5. The control method for an all-terrain adaptive mobile chassis according to claim 4, characterized in that, The step of calculating the expected rotation value of the rotary joint and the expected stroke value of the linear actuator corresponding to the expected relative coordinate position of each wheel-leg composite unit includes: For each of the wheel-leg composite units, the corresponding grounding contact position is extracted based on the desired relative coordinate position to establish a first kinematic geometric constraint condition that reflects the relationship between the swing arm adjustment angle and the corresponding grounding contact position. Establish a second kinematic geometric constraint condition that reflects the relationship between the telescopic adjustment length and the position of the grounding contact; Under the premise of satisfying the first kinematic geometric constraints and the second kinematic geometric constraints, the optimization objective is to minimize the overall energy consumed by the wheel-leg composite unit in performing the swing arm movement and the extension and retraction movement. Then, within the feasible domain of the joint, the expected value of the rotation of the rotary joint and the expected value of the stroke of the linear actuator are calculated.

6. A control device for an all-terrain adaptive mobile chassis, characterized in that, include: The feedback acquisition module is used to acquire the terrain elevation feedback sequence of multiple independently driven wheel-leg composite units on the all-terrain adaptive mobile chassis at the current moment. The feature extraction module is used to perform spatial distribution analysis on the terrain elevation feedback sequence in order to extract the terrain abrupt change features of the current environment; The attitude determination module is used to input the terrain elevation feedback sequence and the terrain abrupt change features into a pre-configured attitude decision network to determine the target vehicle attitude angle of the all-terrain adaptive mobile chassis. The adjustment parameter calculation module is used to calculate the swing arm adjustment angle and telescopic adjustment length of each of the multiple independently driven wheel-leg composite units according to the target vehicle body attitude angle. The motion control module is used to control the corresponding wheel-leg composite unit to perform the swing arm motion according to the swing arm adjustment angle, and after the swing arm motion is completed, control the corresponding wheel-leg composite unit to perform the telescopic motion according to the telescopic adjustment length. The parameter update module is used to iteratively update the network mapping parameters in the attitude decision network based on the chassis balance state characteristics after performing the above actions.

7. An electronic device, characterized in that, include: At least one processor; as well as, A memory that is communicatively connected to the at least one processor; The memory stores a computer program that can be executed by the at least one processor, which, when executed by the at least one processor, causes the at least one processor to perform the control method for the all-terrain adaptive mobile chassis as described in any one of claims 1-9.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program; When the computer program is executed by the processor, it implements the control method for the all-terrain adaptive mobile chassis as described in any one of claims 1-9.

9. A mobile robot, characterized in that, include: The robot itself; An all-terrain adaptive mobile chassis is mounted on the robot body, and the all-terrain adaptive mobile chassis includes multiple independently driven wheel-leg composite units; The terrain elevation feedback acquisition component is used to acquire the terrain elevation feedback sequence of the multiple independently driven wheel-leg composite units at the current moment; The chassis controller is communicatively connected to the terrain elevation feedback acquisition component and the multiple independently driven wheel-leg composite units; The chassis controller is used to perform spatial distribution analysis on the terrain elevation feedback sequence in order to extract the terrain abrupt change features of the current environment. The chassis controller is also used to input the terrain elevation feedback sequence and the terrain abrupt change features into a pre-configured attitude decision network to determine the target vehicle attitude angle of the all-terrain adaptive mobile chassis. The chassis controller is also used to calculate the swing arm adjustment angle and telescopic adjustment length of each of the multiple independently driven wheel-leg composite units according to the target vehicle body attitude angle. The chassis controller is also used to control the corresponding wheel-leg composite unit to perform the swing arm action according to the swing arm adjustment angle, and after the swing arm action is completed, control the corresponding wheel-leg composite unit to perform the telescopic action according to the telescopic adjustment length. The chassis controller is also used to iteratively update the network mapping parameters in the attitude decision network based on the chassis balance state characteristics after performing the above actions.

10. A control system for an all-terrain adaptive mobile chassis, characterized in that, include: An all-terrain adaptive mobile chassis, comprising multiple independently driven wheel-leg composite units; A terrain elevation feedback acquisition device is used to acquire the terrain elevation feedback sequence of the multiple independently driven wheel-leg composite units at the current moment; The attitude control device is communicatively connected to the terrain elevation feedback acquisition device and the multiple independently driven wheel-leg composite units; The attitude control device is used to perform spatial distribution analysis on the terrain elevation feedback sequence in order to extract the terrain abrupt change features of the current environment. The attitude control device is also used to input the terrain elevation feedback sequence and the terrain abrupt change features into a pre-configured attitude decision network to determine the target vehicle attitude angle of the all-terrain adaptive mobile chassis. The attitude control device is also used to calculate the swing arm adjustment angle and telescopic adjustment length of each of the multiple independently driven wheel-leg composite units according to the target vehicle body attitude angle. The attitude control device is also used to control the corresponding wheel-leg composite unit to perform the swing arm action according to the swing arm adjustment angle, and after the swing arm action is completed, control the corresponding wheel-leg composite unit to perform the telescopic action according to the telescopic adjustment length. The attitude control device is also used to iteratively update the network mapping parameters in the attitude decision network based on the chassis balance state characteristics after performing the above actions.

Images