Embodied intelligence-based adaptive rocker arm type carrying robot control method and system

Abstract

The application discloses a kind of based on embodiment intelligence's self-adapting rocker arm type carrying robot control method and system, belong to robot control field, the method includes: by being carried on robot car body and rocker arm mechanism on multi-modal sensor array obtains first perception dataset and second state dataset, extracts historical configuration matrix and exports feedforward configuration instruction to realize feedforward pre-adjustment;During obstacle crossing and load period, generate and execute the non-symmetrical compensation instruction of driving two sides rocker arm to carry out asymmetric telescoping and lifting;In distress period, according to the attribute of physical root node, match the corresponding quasi-physical escape sequence from state transition matrix and execute escape action;In edge control node, running reinforcement learning network module to output original speed instruction vector.The application guarantees the control completeness when facing new terrain, improves the accuracy of compensation action.

Description

Technical Field

[0001] This invention relates to the field of robot control, and in particular to an adaptive rocker arm handling robot control method and system based on embodied intelligence. Background Technology

[0002] Due to their excellent terrain adaptability, rocker-type handling robots have been widely used in complex working conditions such as mine transportation, construction site material handling, port and dock heavy-load transfer, and disaster relief material delivery. These robots typically use a multi-stage rocker-bogie suspension structure to connect multiple drive wheels. Through the active or passive movement of each rocker joint, the chassis attitude is adaptively adjusted, thereby maintaining stable driving on unstructured terrains such as steps, slopes, ditches, and gravel roads.

[0003] However, as industrial applications increasingly demand higher handling efficiency, load capacity, and autonomy, the existing control systems of rocker-arm handling robots have revealed the following systemic technical deficiencies: First, existing control systems generally employ passive attitude adjustment strategies based on proportional-integral-derivative (PID) feedback loops. However, within the time window between the occurrence of attitude deviation and the activation of compensation actions, the robot is in an uncontrolled transient instability state. The heavy load it carries may experience significant dynamic impact loads due to inertial effects, potentially leading to load displacement or damage, or even causing the entire robot to tip over. Currently, the industry lacks a feedforward control mechanism capable of pre-adjusting the rocker arm configuration before the robot actually steps onto the terrain, thus fundamentally eliminating the inherent adjustment lag defect of passive feedback. Second, existing safety control systems for heavy-duty handling robots generally assume that the center of gravity of the external load is constant and located at the geometric center of the robot's handling platform when conducting stability assessments. However, in real industrial handling operations, the center of gravity of the load may be significantly offset due to eccentric placement during initial loading. Existing systems cannot identify the true mass and three-dimensional center of gravity position of the load in real time without relying on pre-calibration. The stability assessment results have a continuously widening deviation from the actual physical state, resulting in the inability to trigger effective anti-tipping compensation measures in a timely manner under high-risk conditions with severe center of gravity offset.

[0004] In summary, there is an urgent need for a control method for rocker-type handling robots that can systematically solve the aforementioned multiple technical defects. This method should have feedforward pre-adjustment capability, online identification of load center of gravity and active anti-tipping capability, accurate diagnosis of the root causes of difficulties and targeted escape capability, deterministic safety assurance capability of reinforcement learning output, and adaptive degraded operation capability in communication interruption scenarios. Summary of the Invention

[0005] One of the objectives of this invention is to provide an adaptive rocker arm handling robot control method and system based on embodied intelligence, so as to solve the problem of lag in posture adjustment caused by passive feedback control in existing rocker arm handling robots.

[0006] This invention is achieved through the following technical solution: an adaptive rocker arm handling robot control method based on embodied intelligence, comprising the following steps: acquiring a first perception dataset and a second state dataset through a multimodal sensor array mounted on the robot body and rocker arm mechanism; converting the first perception dataset into feature vectors, and performing distance metric matching between the feature vectors and a pre-constructed embodied experience memory map; when a match is found, extracting the historical configuration matrix and outputting a feedforward configuration command to achieve feedforward pre-tuning; during obstacle crossing and load loading periods, establishing and solving a system of linear equations for torque balance containing unknown load mass and unknown center of gravity coordinates based on the normal support force matrix of each rocker arm suspension point and the absolute tilt angle matrix of each rocker arm joint, obtaining the three-dimensional center of gravity coordinates of the load, and constructing a dynamic support polygon based on the three-dimensional center of gravity coordinates of the load. The stability margin of the overall center of gravity projection coordinates is calculated. When the stability margin is lower than the stability margin threshold, an asymmetric compensation command is generated and executed to drive the two rocker arms to perform asymmetric extension and retraction and lifting. During the distress period, a physical directed acyclic graph is constructed based on the causal dependencies between various physical quantities in the robot motion system. By calculating the normalized residuals of each node and performing topological sorting and traversal along the directed edges, the physical root cause nodes causing power loss are isolated. According to the attributes of the physical root cause nodes, the corresponding simulated escape gait sequence is matched from the state transition matrix and the escape action is executed. A reinforcement learning network module is run in the edge control node to output the original velocity command vector. The original velocity command vector is projected into the safe and feasible region defined by the motor torque limit and the link yield stress limit to obtain the safe command vector and issue it for execution.

[0007] Furthermore, the first perception dataset includes terrain elevation data and material reflectivity data, and the second state dataset includes servo feedback current of each rocker arm servo motor and current joint angle of each joint.

[0008] Furthermore, the first sensing dataset and the second state dataset are synchronized and marked in time by a unified hardware clock signal to ensure that the environmental sensing data and the mechanism state data used in the same control cycle are aligned in timestamps.

[0009] Furthermore, the embodied experience memory map is constructed offline by the cloud processing node based on the robot's historical operation data, and distributed to the edge control node for local storage in the form of a hash table; the embodied experience memory map uses the feature signature vector of the historical terrain as the index key and the historical optimal configuration parameter corresponding to the historical terrain as the storage value.

[0010] Further, the step of converting the first perception dataset into a feature vector specifically includes: uniformly dividing the spatial area covered by the terrain elevation data into multiple fixed-size three-dimensional voxel grids; performing local plane fitting on the point cloud sampling points contained in each three-dimensional voxel grid to estimate the surface normal vector at each sampling point, and calculating the variance of all surface normal vectors in the three-dimensional voxel grid to obtain the surface normal vector variance of the three-dimensional voxel grid; extracting the corresponding reflection intensity values ​​for the point cloud sampling points contained in each three-dimensional voxel grid and calculating the arithmetic mean to obtain the mean material reflectivity of the three-dimensional voxel grid; and concatenating the surface normal vector variance and the mean material reflectivity of all three-dimensional voxel grids in voxel order to generate a fixed-dimensional terrain signature vector as the feature vector.

[0011] Further, the distance metric matching includes: in the edge control node, using the terrain signature vector as a query key, inputting it into the embodied experience memory graph constructed based on the locality-sensitive hashing algorithm for retrieval; wherein, the embodied experience memory graph contains multiple independent random projection hash functions, each random projection hash function maps the terrain signature vector to an integer bucket number by performing an inner product operation between the terrain signature vector and a random projection vector, adding a random offset, dividing by the quantization slot width parameter, and rounding.

[0012] Furthermore, when a match is found, extracting the historical configuration matrix and outputting the feedforward configuration instruction includes: calculating the proportion of the terrain signature vector and the historical terrain signature vector stored in the embodied experience memory map that have the same bucket number on all random projection hash functions, and using the complement of this proportion as the hash dissimilarity; when the hash dissimilarity is lower than a first threshold, a match is determined, the historical configuration matrix is ​​extracted from the corresponding hash bucket, and the historical configuration matrix is ​​output as the feedforward configuration instruction; when the hash dissimilarity is not lower than the first threshold, the process switches to real-time solution.

[0013] Furthermore, the historical configuration matrix includes the optimal swing arm extension angle value, the target position coordinates of each joint, and the suspension stiffness damping coefficient; the edge control node converts the historical configuration matrix into a pulse width modulation duty cycle bias signal, and injects the bias signal into the servo control loop of each swing arm joint as a feedforward compensation amount.

[0014] Further, the three-dimensional center-of-gravity coordinates of the load are obtained through the following steps: reading the normal support force matrix obtained by force sensors at each wheel contact point or by mapping the servo feedback current through the motor torque constant, and the absolute tilt angle matrix fed back by the absolute position encoder built into each rocker joint; based on the absolute tilt angle matrix, the three-dimensional spatial coordinates of each wheel contact point are calculated through forward kinematics, and a spatial torque balance equation containing the unknown load mass and the unknown load three-dimensional center-of-gravity coordinates is established; wherein, the sum of the torques generated by the normal support forces at each wheel contact point is equal to the sum of the torques generated by the robot's known self-weight and the unknown load under gravity; the product of the unknown load mass and the unknown load three-dimensional center-of-gravity coordinates in the spatial torque balance equation is defined as a component of the load state vector, and the spatial torque balance equation is rewritten as a system of linear torque balance equations about the load state vector; the system of linear torque balance equations is solved using a singular value decomposition algorithm to obtain the optimal load state vector; the load mass component is extracted from the optimal load state vector, and the remaining components are divided by the load mass component to obtain the three-dimensional center-of-gravity coordinates of the load.

[0015] Further, the optimal load state vector is obtained through the following steps: performing singular value decomposition on the coefficient matrix of the torque balance linear equation system to obtain a left orthogonal singular vector matrix, a singular value diagonal matrix, and a right orthogonal singular vector matrix; taking the reciprocal of the singular values ​​in the singular value diagonal matrix whose values ​​are greater than a preset truncation threshold, and setting the singular values ​​whose values ​​are not greater than the preset truncation threshold to zero to obtain a truncated generalized inverse matrix; multiplying the right orthogonal singular vector matrix, the truncated generalized inverse matrix, and the transpose of the left orthogonal singular vector matrix with the right-hand vector of the torque balance linear equation system to obtain the optimal load state vector.

[0016] Further, the generation and execution of the asymmetric compensation command includes: taking the horizontal components of the three-dimensional coordinates of all bearing wheel contact points, obtaining the convex hull of the planar point set formed by the horizontal components, and obtaining a dynamic support polygon; combining the known center of gravity of the robot and the three-dimensional center of gravity coordinates of the load, calculating the vertical projection coordinates of the overall system center of gravity on the horizontal plane, and obtaining the overall center of gravity projection coordinates; calculating the normal distances from the overall center of gravity projection coordinates to the straight lines of each boundary of the dynamic support polygon, and taking the minimum value as the stability margin; when the stability margin is lower than the stability margin threshold, determining the nearest boundary corresponding to the stability margin as the potential overturning axis, and generating a horizontal displacement vector of the center of gravity pointing from the overall center of gravity projection coordinates to the geometric center of the dynamic support polygon; mapping the horizontal displacement vector of the center of gravity to the joint space through the pseudo-inverse of the Jacobian matrix of the rocker arm system, obtaining the asymmetric elongation and lifting angle adjustment increment of each rocker arm, and issuing and executing the asymmetric compensation command.

[0017] Furthermore, when mapping the horizontal displacement vector of the center of gravity to the joint space using the pseudo-inverse of the Jacobian matrix of the rocker arm system, the method further includes: using the redundant degrees of freedom of the rocker arm system, selecting a null space vector in the null space of the Jacobian matrix, and superimposing the null space projection term corresponding to the null space vector into the asymmetric compensation command, so as to constrain the absolute height of the chassis to remain unchanged while satisfying the target of horizontal displacement of the center of gravity.

[0018] Furthermore, the physical root cause node of the power loss due to isolation includes: defining a set of nodes in the physical directed acyclic graph, wherein each node represents an observable physical quantity, and the set of nodes includes at least chassis acceleration nodes, wheel speed nodes, and rocker arm servo current nodes; defining a set of directed edges in the physical directed acyclic graph, wherein each directed edge represents a physical causal dependency from cause to effect; for each node in the physical directed acyclic graph, calculating the theoretical expected value of the physical quantity represented by the node according to the dynamic model, and calculating the normalized residual between the actual measured value of the node and the theoretical expected value; traversing each node along the directed edges in topological sorting, and when the normalized residual of a node exceeds the fault determination threshold, and the normalized residual of all direct parent nodes of the node in the physical directed acyclic graph does not exceed the normal range determination threshold, the node is determined to be the physical root cause node.

[0019] Furthermore, the fault type of the physical root cause node is classified into one of the following: diagonal suspension, unilateral subsidence, or chassis bottoming out; the step of matching the corresponding simulated escape gait sequence from the state transition matrix and executing the escape action according to the attributes of the physical root cause node includes: performing key-value matching in the state transition matrix according to the fault type classification result of the physical root cause node, obtaining the corresponding simulated escape gait sequence parameters, and executing them.

[0020] Furthermore, when the fault type is diagonal suspension, a bottom-finding operation is performed: the brake of the wheel that is still on the ground is locked, and the rocker arm mechanism on the suspended side is instructed to continuously increase the downward stroke at a preset slow speed. During the descent, the servo layer continuously monitors the contact torque derivative of the servo joint on the suspended side at a sampling frequency no less than a preset frequency. When the contact torque derivative exceeds the grounding determination threshold, it is determined that the suspended wheel has contacted the ground, the descent is immediately stopped, and the rocker arm on that side is locked in the current position.

[0021] Furthermore, when the fault type is unilateral subsidence, a creep compaction operation is performed: the rocker arm mechanism on the subsided side is instructed to perform periodic reciprocating lifting and lowering movements in the vertical direction according to a preset frequency and preset amplitude. During the downward pressing phase of each compaction cycle, a compressive force is applied to the soil under the wheel, and the pressure is released during the lifting phase. The local soil reaction force at the wheel on the subsided side is continuously monitored. When the local soil reaction force exceeds a pre-stored slippage critical threshold, the creep compaction operation is stopped and normal drive is restored.

[0022] Furthermore, the control method also includes: when the difference between the actual displacement speed of the chassis and the theoretical linear speed of the wheel continuously exceeds a preset slip threshold, triggering the construction and traversal process of the physical directed acyclic graph; during the execution of the simulated traction gait sequence, when the difference between the actual displacement speed of the chassis and the theoretical linear speed of the wheel falls back below the preset slip threshold and continues to maintain a preset confirmation time, terminating the execution of the simulated traction gait sequence and restoring normal driving control.

[0023] Further, the safety command vector is obtained through the following steps: based on the known motor torque constant and the linkage stress distribution model under the current configuration, the expected peak motor torque and expected linkage force value corresponding to the direct execution of the original speed command vector are calculated; the expected peak motor torque is compared with the pre-stored peak torque limits of each motor, and the expected linkage force value is compared with the pre-stored yield stress limit of the linkage material; if any expected value exceeds the corresponding limit value, then using convex quadratic programming, within the safe feasible region jointly defined by the motor torque limit constraint and the linkage yield stress limit constraint, the feasible command with the smallest weighted norm distance to the original speed command vector is selected as the safety command vector.

[0024] Furthermore, the control method further includes: establishing a dual-buffered verification mechanism independent of the edge control node in the servo layer, wherein the first buffer receives the target contact force from the edge control node, and the second buffer cyclically writes the real-time contact force sampling values ​​of each rocker arm torque sensor; the field-programmable gate array hardware logic inside the servo driver directly compares the difference between the target contact force in the first buffer and the real-time contact force sampling value in the second buffer, and generates a pulse width modulation duty cycle command in a proportional-derivative control manner based on the difference and the time change rate of the difference, driving each servo motor to execute force control closed loop.

[0025] Furthermore, the control method also includes: communication adaptive degradation logic: the edge control node periodically sends heartbeat packets to the cloud processing node; when the heartbeat packets are lost for more than a preset number of cycles, the edge control node determines that the communication link with the cloud processing node is interrupted, automatically blocks the instruction input interface from the cloud processing node, and switches to local island operation mode; in the local island operation mode, the retrieval range of the embodied experience memory map is limited to the most recent preset number of trajectory records in the circular buffer residing in the local memory of the edge control node, and the stability margin threshold is increased by a safety margin constant.

[0026] Another aspect of the present invention provides an adaptive rocker arm handling robot control system based on embodied intelligence, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements any of the adaptive rocker arm handling robot control methods based on embodied intelligence as described above.

[0027] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0028] 1. This invention constructs an embodied experience memory map and uses the Locality Sensitive Hashing (LSH) algorithm to achieve fast approximate nearest neighbor matching of terrain signature vectors. This allows the robot to pre-adjust its rocker arm configuration based on historical experience before actually stepping onto the terrain ahead, thereby transforming the control logic from a traditional passive error correction paradigm to an active preparation paradigm and eliminating the inherent adjustment lag of PID feedback control. At the same time, the LSH algorithm compresses the control computation in repetitive terrain scenarios from solving a system of nonlinear differential equations to a single hash calculation and a single threshold comparison, reducing the computational complexity to an approximate constant level. This makes it possible to achieve real-time feedforward retrieval on edge control nodes with limited computing resources. When a match is not found, the system can automatically switch to the real-time solution process, ensuring control integrity when facing entirely new terrain.

[0029] 2. This invention employs an online identification method based on spatial moment balance equations and singular value decomposition algorithms. Without relying on any pre-calibrated information, it continuously identifies the true mass and three-dimensional center of gravity coordinates of external loads using only high-frequency collected measurements of the ground force of each wheel and the angle of each rocker arm joint. By transforming the nonlinear product relationship between load mass and center of gravity coordinates into a linear equation system of mass-mass moment joint state vectors, and using truncated singular value decomposition to suppress numerical amplification errors caused by ill-conditioned matrices, robustness and accuracy of the identification results are achieved. Furthermore, a dynamic support polygon is constructed, and the stability margin of the overall center of gravity projection coordinates is calculated. When the margin is insufficient, the required horizontal displacement of the center of gravity is mapped to the joint space using the pseudo-inverse of the Jacobian matrix, generating asymmetric compensation commands for both rocker arms. Unlike traditional symmetrical adjustment methods that only change the chassis height, asymmetric adjustment can actually translate the horizontal projection point of the overall center of gravity, thereby achieving effective active anti-overturning control. The introduction of the zero-space projection term further ensures that no additional vertical disturbance is introduced while achieving the horizontal displacement target of the center of gravity, improving the accuracy of the compensation action.

[0030] 3. This invention achieves precise localization of the root cause of power loss by constructing a directed acyclic graph based on physical causal dependencies and using normalized residuals for topological sorting and traversal. This method can clearly distinguish between fundamentally different fault modes such as diagonal suspension, single-sided sinking, and chassis bottoming. It also matches a targeted mimicry gait sequence for each fault mode through deterministic key-value matching of the state transition matrix, eliminating the blind struggle behavior of traditional controllers when facing difficulties from the algorithm mechanism level. At the same time, a deterministic hardware constraint filter based on convex quadratic programming is set between the reinforcement learning network module and the servo execution link. This filter transforms the peak torque limit of the motor and the yield stress limit of the connecting rod material into linear inequality constraints of the safe feasible region of the convex polyhedron. When any reinforcement learning output violates the physical constraints, the original instruction is projected to the nearest feasible point in the safe feasible region by solving the weighted norm distance minimization problem, thus maximizing the optimization effect of the reinforcement learning strategy while ensuring absolute physical safety.

[0031] 4. This invention provides adaptive degradation logic for communication in extreme network segmentation scenarios. When the edge control node detects that the heartbeat packets with the cloud processing node are lost continuously for more than a preset number of cycles, it automatically blocks the input of cloud commands that may be outdated, switches to local island operation mode, and achieves continuous autonomous operation capability in a completely isolated environment with the principle of safety first and performance moderately compromised, by converging the retrieval range of the embodied experience memory map to the most recent finite number of trajectory records in the local circular buffer and increasing the stability margin threshold to obtain a greater anti-overturning safety margin. This ensures the physical safety of the robot under extreme communication conditions. Attached Figure Description

[0032] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0033] Figure 1 The above is a flowchart of the overall method provided in Embodiment 1 of the present invention.

[0034] Figure 2 This is a schematic diagram of the feature distribution of the terrain signature vector provided in Embodiment 1 of the present invention.

[0035] Figure 3 This is a schematic diagram illustrating the effect of the number of hash functions on the collision probability curve provided in Embodiment 1 of the present invention.

[0036] Figure 4 The correlation verification curve is provided for Embodiment 1 of the present invention.

[0037] Figure 5 The image shows a comparison curve of the rocker arm response delay provided in Embodiment 1 of the present invention.

[0038] Figure 6 The graph shows the relationship between the identification error and the number of iterations of the singular value decomposition algorithm provided in Embodiment 1 of the present invention.

[0039] Figure 7 This is a comparison chart of the time-varying curve of the dynamic center of gravity offset during the obstacle crossing process and the identification accuracy of the singular value decomposition algorithm provided in Embodiment 1 of the present invention.

[0040] Figure 8 This is a time-series evolution diagram of the dynamic centroid projection point trajectory and stability margin provided in Embodiment 1 of the present invention.

[0041] Figure 9 This is a diagram illustrating the asymmetry of the left and right rocker arm adjustment amounts provided in Embodiment 1 of the present invention.

[0042] Figure 10 The above is a heat map of node residuals under three types of fault scenarios provided in Embodiment 1 of the present invention.

[0043] Figure 11 This is a waterfall-style timing diagram of the normalized residuals of each node provided in Embodiment 1 of the present invention.

[0044] Figure 12 This is a comparison diagram of the escape process provided in Embodiment 1 of the present invention.

[0045] Figure 13 The real-time monitoring curve and grounding event capture diagram provided in Embodiment 1 of the present invention.

[0046] Figure 14This is a comparison chart of force tracking errors provided in Embodiment 1 of the present invention.

[0047] Figure 15 The curve showing the effect of the stability margin threshold change on the rollover probability provided in Embodiment 1 of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0049] Example 1:

[0050] This embodiment discloses an adaptive rocker arm handling robot control method based on embodied intelligence. Figure 1 The overall method flowchart of this embodiment is shown. As can be seen from the figure, this embodiment includes the following steps:

[0051] Step 1: Obtain the first sensing dataset and the second state dataset through a multimodal sensor array.

[0052] The multimodal sensor array refers to a collection of sensors mounted on the robot body and rocker arm mechanism, which includes the ability to perceive different physical quantities. This sensor array includes at least depth sensors (such as 3D LiDAR, structured light depth cameras, or binocular stereo vision cameras) for acquiring three-dimensional spatial information of the terrain ahead, reflectivity sensors (such as those using LiDAR echo intensity channels or dedicated near-infrared reflectivity sensors) for acquiring information on the surface material properties, and servo internal sensors (such as current sensors, encoders, and tilt sensors) for acquiring information on the motion state of the rocker arm mechanism.

[0053] The first perception dataset refers to the collection of raw perception data acquired by a multimodal sensor array to describe the physical properties of the environment in front of the robot. Specifically, the first perception dataset includes two parts: terrain elevation data and material reflectivity data.

[0054] Among them, the terrain elevation data ahead refers to the three-dimensional point cloud data obtained by the depth sensing sensor scanning the ground area in front of the robot in real time. Each sampling point in the point cloud data contains its three-dimensional spatial coordinates in the world coordinate system or the robot body coordinate system. Therefore, it can completely describe the geometric undulation characteristics of the terrain ahead, including information such as step height, slope inclination angle, ditch depth and gravel distribution.

[0055] Material reflectivity data refers to the information on the intensity of electromagnetic wave reflection from the ground surface received by sensors. Different ground materials exhibit significant differences in their absorption and reflection characteristics of laser or near-infrared light. For example, the reflectivity of soft mud is typically significantly lower than that of hard concrete pavement, while the reflectivity of gravel pavement falls in between. Therefore, material reflectivity data can indirectly characterize the hardness and adhesion properties of the ground surface ahead, and is a key physical parameter influencing the selection of robot motion strategies.

[0056] The second state dataset refers to the data set collected in real time by the internal sensors of the servo motor, which describes the current motion and force state of the rocker arm mechanism. Specifically, the second state dataset includes at least the feedback current values ​​of each rocker arm servo motor and the current absolute angle values ​​of each joint.

[0057] The servo feedback current refers to the real-time operating current value measured and fed back by the driver of the servo motor of each rocker arm joint. This current value can be directly mapped to the motor output torque through the known motor torque constant, thus indirectly reflecting the magnitude of the external load force and internal friction force borne by each rocker arm joint at the current moment.

[0058] The current joint angle refers to the angle value fed back by the absolute position encoder built into each rocker arm joint. This angle value describes the rotational offset of each kinematic pair of the rocker arm mechanism relative to the reference zero position. Through forward kinematics calculation, the current joint angle can be used to calculate the three-dimensional spatial coordinates of each wheel contact point and the attitude of the chassis platform in real time.

[0059] In this embodiment, the depth sensing sensor in the multimodal sensor array can be a 3D LiDAR installed at the front of the robot body, with a horizontal scanning field of view of not less than 120 degrees and a vertical scanning line count of not less than 16 lines, capable of outputting dense 3D point cloud data within a range of at least 10 meters in front of the robot at a rate of not less than 10 frames per second. Current sensors are embedded inside each rocker arm servo driver, with a sampling frequency of not less than 1 kHz. The encoder is a multi-turn absolute position encoder with a resolution of not less than 14 bits, achieving high-precision measurement of joint angles without cumulative error.

[0060] In this embodiment, the first perception dataset and the second state dataset are synchronized and marked in time by a unified hardware clock signal to ensure that the environmental perception data and the mechanism state data used in the same control cycle are strictly aligned in timestamps, so as to avoid false estimation errors caused by time offset in data fusion.

[0061] Understandably, the first perception dataset contains the physical information of the environment that the robot sees in front of it, while the second state dataset contains the internal motion and force information that the robot itself senses. The former provides environmental input for terrain matching and feedforward pre-adjustment in subsequent steps, while the latter provides internal state input for dynamic center of gravity identification and attitude compensation in subsequent steps.

[0062] Step 2: Convert the first perception dataset into feature vectors and match them with the embodied experience memory map using a distance metric to achieve feedforward pre-tuning.

[0063] Among them, the feature vector refers to a fixed-dimensional numerical vector obtained from the first perception dataset after feature extraction processing. This numerical vector encodes the geometric features of the terrain ahead and the surface material properties in a compact mathematical form, which is used for subsequent fast similarity retrieval.

[0064] Embodied experience memory graphs are structured knowledge bases built offline by cloud processing nodes based on historical robot operation data and distributed to edge control nodes for local storage in the form of compact hash tables. These memory graphs use the feature signature vectors of historical terrain as index keys and the corresponding historical optimal robot configuration parameters (including arm span angle, target positions of each joint, and suspension damping coefficients) as storage values. Essentially, they constitute an efficient mapping relationship library from terrain features to optimal configurations.

[0065] Understandably, the core technical objective of this step is to solve the inherent adjustment lag problem of traditional PID passive feedback control. That is, traditional control systems must wait until the robot has stepped onto complex terrain and generated posture deviation before triggering compensation actions. Through the matching mechanism of embodied experience memory map, the robot can complete the pre-adjustment of the rocker arm configuration based on historical experience before actually stepping onto the terrain ahead, fundamentally changing the control logic from passive error correction to active preparation.

[0066] In this embodiment, this step is performed before driving or during a stable driving period, that is, the configuration preparation is completed in advance before the robot enters a high-difficulty obstacle-crossing situation.

[0067] In this embodiment, the first perceptual dataset is converted into a feature vector and matched with the embodied experience memory map using a distance metric to achieve feedforward pre-tuning. Specifically, this includes the following sub-steps:

[0068] Sub-step 1: Divide the terrain elevation data ahead into a fixed-size three-dimensional voxel mesh, extract the surface normal vector variance and the mean material reflectivity of the mesh, and cascade them to generate a fixed-dimensional terrain signature vector.

[0069] In this context, a 3D voxel mesh refers to uniformly dividing the spatial region covered by the 3D point cloud data of the foreground terrain into multiple fixed-size cubic units, each cubic unit being called a voxel. The total number of voxel meshes is denoted as... This number is determined by both the spatial extent of the point cloud coverage and the side length of a single voxel. The purpose of voxelization is to transform the spatially unevenly distributed raw point cloud data into a structured, regular grid representation, facilitating subsequent independent feature extraction for each local region.

[0070] The variance of the surface normal vector refers to the variance of the first normal vector. For all point cloud sampling points within a voxel, the surface normal vector at each sampling point is first estimated using local plane fitting (e.g., principal component analysis). Then, the statistical variance of all normal vectors within the voxel is calculated. The normal vector describes the orientation of the local surface, while the variance of the normal vector directly reflects the geometric roughness of the terrain within the voxel. A larger variance indicates more severe terrain undulations and more irregular surface orientation changes within the local area. The robot's suspension system needs to handle greater impact loads when traversing this area. The variance of the surface normal vector of an individual element is denoted as .

[0071] The average reflectance of a material refers to the reflectance of the first... For each point cloud sampling point within a voxel, the laser echo reflection intensity value corresponding to each sampling point is extracted, and the arithmetic mean of all reflection intensity values ​​within that voxel is calculated. This mean is denoted as . The physical significance of this is that different surface materials (such as soil, sand, concrete, and metal plates) have significant differences in their reflectivity to lasers, and the average reflectivity can be used as an indirect indicator to distinguish the hardness and friction characteristics of the surface.

[0072] In this embodiment, the process of constructing the terrain signature vector is as follows: All The surface normal variance of each individual element and the average reflectivity of the material By concatenating the voxels in sequence, a unified real-valued vector is formed.

[0073] For example, in this embodiment, the terrain signature vector can be constructed using the following formula:

[0074]

[0075] in, The vector order concatenation operator represents the concatenation of the two-dimensional subvectors corresponding to each voxel according to their voxel indices from the 1st to the 2nd voxel. One after another, they are connected end to end; For the first The variance of the surface normal vector of an individual prime point set is used to quantify the geometric roughness of that local region. For the first The mean laser reflectance of an individual point set is used to reflect the hardness or softness of the surface material. This is the transpose symbol for a matrix; The total number of voxel grids determines the dimension of the signature vector and the spatial resolution of the terrain description; The final generated current terrain signature vector has dimensions of It simultaneously encodes information in two dimensions: the geometric shape and the physical material of the terrain.

[0076] Understandably, by concatenating and encoding geometric roughness features and material hardness features in the same vector, the terrain signature vector can comprehensively describe the physical properties of a terrain segment in a compact mathematical form. The subsequent hash matching process will determine the similarity between two terrain segments based on the distance between these signature vectors in high-dimensional space. Figure 2 This diagram illustrates the feature distribution of the terrain signature vector in this embodiment. Figure 2 The study demonstrates the cluster distribution of different typical terrains in the feature space of normal vector variance-reflectivity mean, distinguishing five typical terrain categories (concrete flats, gravel, soft mud, wooden steps, and metal ramps). The signature vectors of each terrain category form clearly separable clusters, with the inter-class distance being significantly greater than the intra-class distance, reflecting the discriminative power of terrain signatures.

[0077] Sub-step 2 involves using the terrain signature vector as a query key in the edge control node and inputting it into the embodied experience memory graph constructed based on the locality-sensitive hashing algorithm for retrieval.

[0078] Edge control nodes are embedded computing devices deployed on the robot itself, responsible for executing control decision-making tasks with high real-time requirements and moderate computational complexity. In terms of computing power, edge control nodes fall between cloud servers and underlying servo drives, enabling them to complete medium-scale computational tasks locally, including hash retrieval, kinematics solving, and reinforcement learning inference, without relying on real-time network communication with cloud servers.

[0079] Locality-Sensitive Hashing (LSH) algorithms are a class of randomized hashing methods specifically designed for near-nearest neighbor searches in high-dimensional spaces. Their core characteristic is that two vectors with a small Euclidean distance in the original high-dimensional space are significantly more likely to fall into the same hash bucket after being mapped by the LSH function than two vectors with a large Euclidean distance. This property allows for the transformation of computationally expensive high-dimensional exhaustive searches (whose time complexity is linearly proportional to the number of records in memory) into low-time-complexity hash bucket lookups, which is crucial for real-time retrieval on edge hardware.

[0080] In this embodiment, the LSH function family in the embodied experience memory map includes An independent random projection hash function Each hash function maps the input high-dimensional signature vector to an integer bucket number.

[0081] For example, in this embodiment, each LSH hash function can be represented by the following formula:

[0082]

[0083] in, For the first The random projection vector used by the hash function, each component of which is independently sampled from a standard normal distribution, is used to project the high-dimensional signature vector onto a one-dimensional real axis. In the interval The random offset of uniform sampling is used to eliminate the bias effect of quantization boundary and prevent the signature vector located near the quantization boundary from being systematically misassigned. The slot width parameter controls the granularity of the hash buckets. The larger the slot width, the wider the signature space covered by each hash bucket, the looser the matching conditions, and the higher the recall rate but the lower the precision; conversely, the matching is more stringent. This is a floor function that discretizes the continuous projection results into integer bucket numbers. Figure 3 This embodiment illustrates the impact of the number of hash functions on the collision probability curve at different real distances. Figure 3 Multiple curves are plotted, each representing a signature vector pair with different true Euclidean distances. It can be seen that as m increases, the collision probability of close-range vector pairs remains high, while the collision probability of far-range vector pairs decreases sharply.

[0084] It is understandable that this is achieved through inner product operations. The high-dimensional terrain signature vector is projected onto a scalar value on a one-dimensional real axis. Then, through three steps—adding an offset, dividing by the slot width, and rounding down—this scalar value is quantized into an integer bucket number. Two signature vectors that are close in distance in the original high-dimensional space tend to have similar one-dimensional projection values, thus having a higher probability of falling into the same integer bucket after the same quantization process. This probabilistic distance-preserving property is the mathematical basis for the LSH algorithm's ability to achieve efficient approximate nearest neighbor retrieval.

[0085] Sub-step 3: Determine the matching degree between the current terrain and the historical terrain based on the hash collision ratio. If a match is found, extract the historical configuration matrix and output the feedforward instruction.

[0086] After completing all After mapping using a hash function, the current terrain signature is statistically analyzed. A historical terrain signature stored in the memory map In all The proportion of collisions (i.e., identical hash values) occurring on a hash function can define the hash dissimilarity between two hash functions. A higher collision proportion indicates that the two signature vectors are assigned to the same bucket in multiple independent random projection directions. Therefore, the probability that the two are close in the original high-dimensional space is greater, meaning that the physical properties of the two terrain segments are more similar.

[0087] For example, in this embodiment, hash dissimilarity can be calculated using the following formula:

[0088]

[0089] in, This is an indicator function; it takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. and The current terrain signature and historical terrain signature are respectively... Integer bucket numbers obtained after mapping by each hash function; summation items The calculation is in all The higher the proportion of the two hash functions where the bucket numbers are completely identical, the more similar the two terrain segments are. , which is the hash dissimilarity, is the complement of the collision ratio. The smaller the value, the higher the similarity between the two terrain segments. Figure 4 The graph showing the correlation verification between hash dissimilarity and true Euclidean distance in this embodiment is illustrated. Figure 4 A large number of randomly sampled signature pairs are displayed in the form of scatter plots, and linear fitting trend lines and confidence intervals are superimposed to show the recall and precision within the hit area. This verifies that LSH hash dissimilarity is highly positively correlated with real terrain similarity, and proves the approximate reliability of using hash dissimilarity instead of precise Euclidean distance as a matching criterion.

[0090] In this embodiment, a dissimilarity threshold is set. This threshold, as the first threshold in the claims, is used to determine whether the current terrain is sufficiently similar to historical terrain. When the hash dissimilarity is lower than this threshold, it is determined that the current terrain has appeared in historical experience, and the system skips the real-time trajectory calculation process and directly extracts the corresponding historical optimal configuration parameters from the matched hash bucket.

[0091] For example, in this embodiment, the feedforward triggering logic can be represented by the following formula:

[0092]

[0093] in, The dissimilarity threshold (i.e., the first threshold) is used. When the dissimilarity is lower than this threshold, the current terrain is considered to be sufficiently similar to the historical terrain, and the historical best configuration can be directly reused. This is the historical configuration matrix extracted from the corresponding hash bucket. This matrix contains complete feedforward configuration parameters such as the optimal swing arm angle value, the target position coordinates of each joint, and the suspension stiffness and damping coefficient. This is the output feedforward configuration instruction.

[0094] In this embodiment, the edge control node converts the historical configuration matrix into a pulse width modulation (PWM) duty cycle bias signal. This bias signal is directly injected into the servo control loop of each rocker arm joint as a feedforward compensation amount, which is used to adjust the extension angle and suspension stiffness of the rocker arm mechanism in advance without waiting for the feedback error signal to be established, thereby achieving the physical effect of pre-adjustment.

[0095] It is understandable that the decision branches of the feedforward matching mechanism have completely deterministic characteristics: It is a clear Boolean condition that does not involve any random sampling or probabilistic inference. Once this condition is met, the system completely bypasses the complex real-time kinematic differential equation solving process, directly reading the answer from the historical best configuration library and outputting it. This design compresses the control computation in repetitive terrain scenarios from solving a system of nonlinear differential equations to a single hash calculation and a single threshold comparison, fundamentally eliminating the inherent adjustment lag of traditional passive feedback at the algorithmic mechanism level. When a match is not found (i.e....), When the current terrain is a new terrain that has never appeared in historical experience, the system proceeds to step S300 to perform real-time dynamic center of gravity identification and active attitude compensation. Figure 5 The graph shows a comparison of the rocker arm response delay between feedforward pre-tuning and traditional PID passive feedback in this embodiment. Figure 5The X-axis represents time, and the Y-axis represents the rocker arm extension angle. The two curves represent the delayed response of traditional PID, which starts adjusting only after stepping onto the terrain, and the advanced response of the feedforward matching mechanism in this embodiment, which completes pre-adjustment before stepping onto the terrain. It can be seen that the feedforward matching mechanism in this embodiment can complete the rocker arm configuration adjustment about 0.5 to 1 second in advance, fundamentally eliminating the inherent adjustment lag of traditional passive feedback.

[0096] Step 3: During the obstacle crossing and load-bearing phases, the total mass and three-dimensional center of gravity offset of the dynamic load are calculated in real time, and the two rocker arms are driven to perform asymmetrical extension and retraction and lifting operations.

[0097] The obstacle-crossing and load-bearing phase refers to the stage where the robot is performing challenging maneuvers such as traversing steps, ramps, and ditches, while carrying industrial loads (such as steel coils, building materials, and liquid containers). During this phase, the true center of gravity of the load may be significantly offset from the initial moment due to eccentric placement, and will continue to shift in three-dimensional nonlinearity as the robot's various arm configurations dynamically change during obstacle-crossing.

[0098] Dynamic loads refer to external load objects placed on a robot handling platform whose mass and center of gravity shift dynamically during operation due to changes in the robot's posture and road surface excitation.

[0099] Total mass refers to the sum of the robot chassis's own weight and the mass of the external dynamic load. This total mass is a fundamental parameter for calculating the overall system's center of gravity and evaluating its stability.

[0100] The three-dimensional center of gravity offset refers to the spatial offset distance of the true three-dimensional center of gravity coordinates of the external dynamic load relative to the geometric center of the robot platform. This offset includes components in three directions: the longitudinal axis (front and back), the transverse axis (left and right), and the vertical axis (up and down).

[0101] Understandably, the problem this step aims to solve is that existing safety control systems for heavy-duty handling robots generally assume the load's center of gravity is fixed at the geometric center of the robot platform. However, in actual heavy-duty operation scenarios, the true position of the load's center of gravity may be offset not only during initial loading but also continuously shifts with the dynamic changes in the rocker arm configuration. When this shift causes the vertical projection point of the overall system's center of gravity to approach or even cross the boundary of the support polygon formed by the wheel contact points, the risk of rollover increases dramatically. Therefore, it is essential to first identify the load's mass and three-dimensional center of gravity position in real time without relying on pre-calibration before accurately calculating how to asymmetrically adjust the two rocker arms to ensure system stability. This can specifically include the following sub-steps:

[0102] Sub-step 1: Read the normal support force matrix of each rocker arm suspension point and the absolute tilt angle matrix of each rocker arm joint.

[0103] Among them, the normal support force matrix This refers to a matrix composed of the normal grounding force values ​​at each grounding point, obtained from torque sensors at each wheel hub or by mapping the feedback current from the servo motor through the motor torque constant. For a six-wheel rocker arm-bogie structure, this matrix contains the normal grounding force components of each of the six grounding points.

[0104] Absolute tilt matrix This refers to the vector or matrix composed of the current joint angle values ​​fed back by the absolute position encoders built into each rocker arm joint. Through forward kinematics calculations, this tilt matrix can be used to determine the three-dimensional spatial coordinate position of each wheel contact point in the world coordinate system.

[0105] In this embodiment, the normal support force matrix is ​​read when the robot is stationary or moving in a uniform straight-line state to satisfy the quasi-static assumptions required by the subsequent moment balance equations. When the robot is accelerating, the normal support force matrix can be corrected by introducing an inertial force compensation term to extend the applicability of this method to non-static working conditions.

[0106] Sub-step 2: Establish a system of linear equations for torque balance that includes the unknown load mass and the unknown center of gravity coordinates.

[0107] The key physical insight for solving the load identification problem lies in the fact that when the robot is statically or quasi-statically supported on the ground, the normal support forces at the six wheel contact points must satisfy the three-dimensional moment balance equations in space, together with the robot's own weight and the load's gravity. Therefore, if the normal ground force of each wheel can be measured in real time, and the three-dimensional coordinates of each wheel contact point can be obtained through forward kinematics calculations using the current rocker arm configuration, the originally unknown load mass and three-dimensional center of gravity position can be transformed into a problem of solving a structured linear equation system.

[0108] Specifically, for a six-wheeled rocker robot, under the quasi-static assumption, the sum of the torques of all external forces about any reference point is zero.

[0109] For example, in this embodiment, the three-dimensional moment balance equation in space can be expressed as follows:

[0110]

[0111] in, Based on the current rocker arm joint angle The first obtained through forward kinematics calculation The three-dimensional coordinate vector of each wheel contact point is updated in real time as the rocker arm configuration changes; For the first The normal grounding force vector of each grounding point is obtained by the hub torque sensor or current mapping. The known weight of the robot chassis is calibrated at the factory and stored in the system parameters. The known center of gravity position of the chassis itself; The external load mass to be identified is unknown; Let be the three-dimensional centroid coordinate vector of the load to be identified, and be the unknown quantity; This is the vector of gravitational acceleration; This is a vector cross product operation used to calculate torque.

[0112] It is understandable that the left side of the above equation contains a total of , , , There are four unknowns. By algebraically rearranging the three-dimensional moment balance equations (containing three scalar equation components) according to the unknowns, they can be rewritten as a function of the state vector of the unknown load. linear equations in form The coefficient matrix The elements are determined by the current rocker arm configuration. The right-hand vector is determined by the coordinates of each grounding point. It is determined by the currently measured ground forces and the known chassis self-weight moment term. Using the above-described joint mass-mass moment state vector definition method, the original state vector concerning... and The nonlinear product relationship is linearized, making the system of equations solvable using standard linear algebra methods.

[0113] Sub-step 3: The singular value decomposition (SVD) algorithm is used to solve the linear equation system and extract the three-dimensional centroid coordinates of the load at the current moment.

[0114] In real-world operating scenarios, the presence of sensor measurement noise and the over-constrained structure of the rocker arm may lead to issues with the coefficient matrix. If the condition number is large or even close to singular (i.e., there are singular values ​​close to zero in the matrix), directly inverting the coefficient matrix to solve the system of equations will drastically amplify the numerical error of the solution, rendering it impractical. Therefore, the robust least squares solution of this system of equations is obtained using the singular value decomposition method. This is a deterministic numerical solution algorithm that does not involve any random sampling process.

[0115] For example, in this embodiment, the least-squares identification solution of the load state vector can be obtained by the following formula:

[0116]

[0117] in, Coefficient matrix The singular value decomposition results, It is a left orthogonal singular vector matrix. It is a right orthogonal singular vector matrix. It is a singular value diagonal matrix; The truncated generalized inverse of the singular value matrix is ​​obtained by: [The operation is as follows:] Non-zero singular values ​​in the matrix are taken in reverse order. For singular values ​​that are close to zero (below a preset threshold), they are set to zero directly instead of taking their reciprocals, thereby suppressing the numerical amplification error of the solution caused by the ill-conditioned matrix. The optimal load state vector obtained is used for identification.

[0118] In this embodiment, from the optimal load state vector The method for separating the load mass and three-dimensional centroid coordinates is as follows: take the fourth component as the identified load mass. Then divide the first three components by respectively The three-dimensional centroid coordinates of the load can then be obtained. The results are used to refresh the dynamic constraint boundary parameters in the edge control nodes. Figure 6 The graphs showing the relationship between the identification error and the number of iterations of the singular value decomposition algorithm under different load eccentricities in this embodiment are illustrated. Figure 6 The convergence speed curves of the SVD identification algorithm under different load eccentricities (light load eccentricity, heavy load light eccentricity, heavy load heavy eccentricity) are shown. All curves should converge to millimeter-level accuracy within a finite number of steps. It can be seen that the second-level singular value decomposition algorithm identification method can accurately identify different degrees of load eccentricity within a short observation window, i.e., within a small number of sensor frames, without the need for pre-calibration. Figure 7 The diagram shows a comparison between the time-varying curve of the dynamic center of gravity offset during the obstacle crossing process and the identification accuracy of the singular value decomposition algorithm in this embodiment. Figure 7 The three curves in the figure are: the actual center of gravity trajectory, the real-time identification result of SVD, and the traditional fixed center of gravity assumption. It can be seen that the SVD identification value closely tracks the actual value, while the deviation of the fixed assumption continues to increase as the obstacle crossing progresses. It can track the three-dimensional displacement of the load center of gravity in real time with millimeter-level accuracy without relying on pre-calibration, while the error of the traditional fixed assumption can reach tens of millimeters.

[0119] Understandably, the above identification process can continuously identify the three-dimensional center of gravity offset of the pure load using only high-frequency sensing data (grounding force measurement value and joint angle measurement value) without any prior calibration, providing real-time updated dynamic input for the next step of asymmetric compensation calculation.

[0120] Sub-step 4: Construct a dynamic support polygon, calculate the stability margin of the overall centroid projection coordinates, and generate an asymmetric compensation command when the margin is insufficient.

[0121] After obtaining the three-dimensional centroid identification results of the load in real time, the next step is to assess whether the current overall system centroid is within a stable range, and to calculate the precise active compensation amount when the stability margin is insufficient.

[0122] Among them, the dynamic support polygon refers to taking the coordinate vector of the ground contact points of all currently bearing wheels. The set of planar points formed by the horizontal components The convex polygon obtained by finding its convex hull is the robot's current stable support region. According to the stability criterion of the support polygon, the robot is in a statically stable state when the vertical projection point of the robot's center of gravity falls inside the convex polygon; once the vertical projection point of the center of gravity exceeds the boundary of the convex polygon, rollover will be inevitable.

[0123] The overall center of gravity projection coordinates refer to the vertical projection coordinates of the overall system center of gravity on the horizontal plane after combining the known self-weight center of gravity of the chassis and the real-time identified load center of gravity. .

[0124] Stability margin refers to the minimum normal distance from the overall center of gravity projection coordinates to the boundary lines of each boundary line of the dynamic support polygon (convex hull). The larger this distance value, the farther the system is from the rollover boundary and the better its stability.

[0125] For example, in this embodiment, the minimum stability margin from the center of gravity to the boundary of the supporting convex hull can be calculated by the following formula:

[0126]

[0127] in, It is the set of convex hulls of the horizontal projections of the six grounding points; The first convex hull The boundary line segment has the following equation: , , , These are the standard equation coefficients of the line; The horizontal projection coordinates of the center of gravity of the entire system; It is the normal distance from the current global dynamic centroid projection point to the nearest boundary among all convex hull boundary lines, which is the current minimum stability margin. Figure 8 The following diagram illustrates the time-series evolution of the dynamic centroid projection point trajectory and stability margin in this embodiment. Figure 8 (a) is a top view showing the positions of the six grounding wheels, the boundary of the supporting convex hull, the trajectory of the center of gravity projection point without compensation, and the trajectory of the center of gravity projection point with compensation. Figure 8 (b) is a time series plot, with the X-axis representing time and the Y-axis representing the minimum stability margin. Display security threshold The comparison of the horizontal line and the time series of margin in the two cases with and without compensation shows that during the heavy-load obstacle crossing process, the center of gravity projection point goes out of the safe domain (margin falls below the threshold) in the case of no compensation, while the asymmetric Jacobian compensation forces it to stay within the safe domain, eliminating the risk of rollover.

[0128] In this embodiment, a safety margin threshold is set. That is, the stability margin threshold in the claims. Below When this occurs, it indicates that the center of gravity projection point is too close to the support boundary, meaning that the condition that the Euclidean distance between the projection coordinates in the claims and the geometric center of the support polygon is greater than the stability margin threshold is met, and active compensation needs to be triggered immediately.

[0129] The goal of compensation is to forcibly translate the horizontal projection point of the overall center of gravity along the direction away from the nearest boundary by asymmetrically adjusting the joint angles of the two rocker arms until the stability margin requirement is met.

[0130] It is understandable that asymmetrical adjustment means that the adjustment applied by the left and right rocker arms is not equal, and the difference is exactly what is necessary to move the center of gravity horizontally—this is fundamentally different from symmetrical adjustment (which can only change the chassis height but cannot move the horizontal projection point of the center of gravity).

[0131] Specifically, the minimum normal distance is first selected. Corresponding convex hull boundary As the current potential overturning axis, determine the unit normal vector pointing from the global centroid projection coordinates to the geometric center of the convex hull. Then generate the required horizontal displacement vector of the center of gravity. ,in The minimum displacement required is then determined. Subsequently, the required horizontal displacement of the center of gravity projection point is mapped to the adjustment increment in the joint space using the pseudo-inverse of the Jacobian matrix of the rocker arm system.

[0132] For example, in this embodiment, the asymmetric adjustment increment of the two rocker arms can be obtained by solving the following formula:

[0133]

[0134] in, This is the Jacobian matrix of the rocker arm system in the current configuration. This matrix establishes a linear mapping relationship between small changes in joint angles and small displacements in the horizontal projected coordinates of the center of gravity. The Moore-Penrose pseudo-inverse of the Jacobian matrix is ​​used to calculate the desired centroidal plane displacement. The reverse mapping is the minimum norm solution of the joint angle adjustment (i.e., selecting the set with the smallest joint adjustment amplitude among all feasible solutions that satisfy the displacement target). Adjust the incremental amounts of asymmetric elongation and lifting angle of each joint of the left rocker arm; Adjust the incremental values ​​for the asymmetric elongation and lifting angle of each joint of the right rocker arm; the difference between the two determines the horizontal translation and direction of the center of gravity. For the null space projection term, where It is the identity matrix. This is an optional null space vector—this term utilizes the redundant degrees of freedom of the rocker arm system, by selecting an appropriate... This can ensure that the target horizontal displacement of the center of gravity is met while simultaneously constraining the absolute height of the chassis to remain unchanged, thereby avoiding additional vertical disturbances introduced by the center of gravity compensation action. Figure 9 This diagram illustrates the asymmetry of the left and right rocker arm adjustment amounts in this embodiment. Figure 9 In the diagram, the X-axis represents the simulation time step, and the Y-axis represents the rocker arm adjustment increment. A bar chart is used to compare and display the adjustment amount of the left rocker arm in the same time series. Adjustment amount with right rocker arm The two have asymmetrical characteristics with unequal values ​​and possibly opposite directions. At the same time, the contribution ratio of the zero-space projection term is marked. It can be seen that the horizontal translation of the center of gravity can only be achieved by the two rocker arms executing differentiated commands. Symmetrical adjustment can only change the chassis height, which verifies the key role of the zero-space projection term in maintaining a constant chassis height.

[0135] In this embodiment, the asymmetric adjustment increment and Decomposed into their respective corresponding asymmetric elongations , and lifting angle , This is used as the step input command for the position loop controller, driving the rocker arm mechanisms on both sides to perform differentiated extension and lifting actions.

[0136] Step 4: During the distress and entrapment period, construct a physical directed acyclic graph and isolate the physical root cause nodes that cause the lack of power.

[0137] The distress period refers to the operational phase in which the robot encounters extreme difficulties during its operation, such as wheel slippage (loss of effective adhesion between the wheel and the ground, the wheel spinning freely while the chassis does not move forward) or partial wheel suspension (a wheel at the end of a rocker arm leaving the ground).

[0138] A directed acyclic graph (DAG) in physics refers to a directed acyclic graph constructed based on the causal dependencies between physical quantities in a robot motion system, with the following structure: .

[0139] A physical root cause node is a node in a directed acyclic graph of physics that is identified as the fundamental cause of the current power loss. The physical link represented by this root cause node is the source of the fault.

[0140] Understandably, the core problem this step addresses is that, faced with wheel slippage or suspension, traditional controllers observe that the chassis movement speed is consistently lower than the commanded speed, thus continuously increasing the motor torque output. However, this torque increase only causes slipping wheels to dig into the ground more quickly, while for suspended wheels, it's completely ineffective, resulting in the robot sinking deeper and deeper into the ground. This step uses physical causal reasoning to accurately pinpoint the root cause of the fault, providing accurate diagnostic evidence for selecting the correct escape strategy in subsequent step 5, thus eliminating blind struggling behavior at the algorithmic mechanism level.

[0141] In this embodiment, when the difference between the actual displacement speed of the chassis and the theoretical linear speed of the wheel continuously exceeds a preset slip threshold, the construction and traversal process of the physical directed acyclic graph is triggered. The actual displacement speed of the chassis can be obtained by measuring the inertial measurement unit (IMU), the theoretical linear speed of the wheel can be calculated by multiplying the rotational speed measured by the wheel encoder by the wheel radius, and the preset slip threshold is a pre-set reference value used to determine whether significant slippage has occurred between the wheel and the ground.

[0142] In this embodiment, constructing a physical directed acyclic graph and isolating physical root cause nodes may specifically include the following sub-steps:

[0143] Sub-step 1: Define the set of nodes and the set of directed edges of the physical directed acyclic graph.

[0144] Among them, the node set Each node in the table represents an observable physical quantity in the robot's motion system. Specifically, the node set includes: chassis acceleration nodes. (The actual measured values ​​are obtained from the inertial measurement unit (IMU), the wheel speed nodes of each wheel) (The actual measurements are obtained from the encoders of each wheel.) (For wheel index), each rocker arm servo current node (The measured values ​​are obtained from each servo motor driver) and the terrain subsidence node (the measured values ​​can be obtained by the depth sensor from the change in the ground elevation of the local area where the wheel is located).

[0145] Directed edge set Each directed edge in the equation represents a physical causal dependency from cause to effect, and this causal direction is determined based on the fundamental laws of rigid body dynamics. For example, the motor current is the direct cause of the wheel hub driving torque, the driving torque is the direct cause of the change in wheel angular velocity, and the wheel angular velocity, in turn, affects the chassis linear acceleration through the ground adhesion force. Therefore, the direction of the directed edge is... .

[0146] Sub-step 2: Calculate in real time the absolute value of the deviation between the measured value of each node and the theoretical expected value based on the dynamic model.

[0147] For each node in the directed acyclic graph Based on the robot's dynamics model and the currently known input conditions, the theoretical expected value of the physical quantity represented by this node under normal operating conditions is calculated. Then, the actual measured value of this physical quantity collected by the sensor is compared with the above theoretical expected value, and the normalized residual between the two is calculated. The larger the residual value, the more serious the deviation of the physical quantity from its normal behavior.

[0148] Sub-step 3: Perform a topological sort traversal along the directed edges to determine the physical root cause node.

[0149] The core logic of root cause localization is: if the normalized residual of a node exceeds the normal range (indicating that the physical quantity represented by the node is in an abnormal state), while the normalized residuals of all its direct parent nodes in the cause-effect graph (i.e., the physical prerequisites or upstream causes of the abnormal quantity) are within the normal range, it means that the abnormality is not transmitted from the upstream physical cause, but originates from an independent failure in the physical link represented by the node itself.

[0150] For example, the Boolean logic for determining the physical root cause in this embodiment can be expressed as follows:

[0151]

[0152]

[0153] in, For nodes The normalized residual is used to measure the degree to which the physical quantity deviates from the theoretical expected value; This is the fault determination threshold. When the normalized residual of a node exceeds this threshold, the physical quantity corresponding to that node is considered to be in an abnormal state. The threshold for determining the normal range is defined as follows: when the normalized residual of a node does not exceed this threshold, the physical quantity corresponding to that node is considered to be within the normal working range. In a directed causal graph Middle node The set of all direct parent nodes, representing All possible upstream physical causes of the anomaly; For logical AND operator. Figure 10 The following diagram shows the node residual heatmaps for three types of fault scenarios used in the physical root cause determination in this embodiment. Figure 10 The X-axis represents the DAG node type, and the Y-axis represents three fault scenarios. The normalized residual of each node is represented by the color intensity of the heatmap. Figure 11 This diagram shows a waterfall-style time series plot of the normalized residuals at each node in the physical root cause determination in this embodiment. Figure 11 The X-axis represents time (seconds), and the Y-axis represents the normalized residual. Figure 11 In the middle (a), the servo current node is shown. Figure 11 (b) represents the wheel speed node. Figure 11 (c) represents the chassis acceleration node. The time of the fault is marked in the figure, showing the typical characteristic of the current node residual being normal while the wheel speed node residual suddenly exceeds the threshold.

[0154] Understandably, let's take a wheel-lifting fault as an example: the typical characteristic of this fault is the servo current node. Normal (motor driver is working properly and providing sufficient current), but the corresponding wheel speed node... The actual value is much higher than the theoretical expected value (the wheel is spinning at high speed in the air, with no effective ground resistance to constrain its rotational speed). At this time, and Therefore, according to the above Boolean logic, The root cause was identified as a wheel being suspended in mid-air. This diagnosis immediately triggered the system to cut off any further torque increase commands to the suspended wheel, preventing the motor from overheating and causing dents in the ground. Figure 12 The diagram shows a comparison of the escape process using root cause isolation in this embodiment and the traditional torque saturation strategy. Figure 12 The X-axis represents the time (seconds) for attempting to escape, the Y-axis represents the motor output torque (N·m), and the Y-axis represents the depth of the robot chassis sinking (mm, indicating the degree of ground digging). It can be seen that the torque curve of the traditional strategy continuously increases to saturation, and the chassis continues to sink. However, the curve of this solution immediately reduces the torque and switches the strategy after isolation, and the chassis sinking stops and gradually recovers.

[0155] In this embodiment, the root cause node's fault type is specifically classified into one of the following three categories: diagonal suspension (meaning that both wheels on opposite sides of the robot simultaneously leave the ground), unilateral sinking (meaning that one wheel on one side of the robot sinks into soft ground and loses effective traction), or chassis bottoming (meaning that the robot's chassis rests directly on a ground protrusion, causing multiple wheels to simultaneously lose effective ground contact pressure). The above classification result will directly determine which gait sequence to select in the subsequent step 5.

[0156] Step 5: Based on the attributes of the physical root cause node, match the corresponding mimicry escape gait sequence and execute self-rescue actions.

[0157] Among them, the mimicry gait sequence refers to a series of structured robot arm motion choreography schemes pre-stored in the built-in state transition matrix, designed for different types of physical fault root causes. The term mimicry means that the design inspiration of this motion scheme comes from the simulation of the self-rescue behavior of biological or mechanical systems in distress; for example, simulating the wriggling motion of an animal stuck in mud, or simulating the downward probing motion of a probe slowly and tentatively touching unknown ground.

[0158] The state transition matrix is ​​a deterministic lookup table that uses fault type classification labels as input indices and corresponding escape gait sequence parameters as output values. Based on the physical root cause node classification results output in step 4, the edge control node performs key-value matching in the state transition matrix to obtain the complete escape action parameter sequence corresponding to the current fault type.

[0159] In this embodiment, executing the simulated escape gait sequence specifically includes the following sub-cases:

[0160] Sub-case 1: If the physical root cause node is classified as diagonally suspended, perform a bottom-finding operation.

[0161] When a physical root cause node is classified as diagonally suspended, the system first locks the brakes on the wheel currently on the ground to prevent the robot from undergoing unexpected displacement or rotation due to the active rotation of some wheels during the escape operation. Subsequently, the system instructs the rocker arm mechanism on the suspended side to perform a bottom-seeking operation, specifically: continuously increasing the downward stroke of the rocker arm on the suspended side at a constant, slow speed (such as a preset displacement rate in millimeters per second), so that the suspended wheel gradually approaches the ground.

[0162] During the descent, the servo layer continuously monitors the contact torque derivative of the servo joint on that side at a frequency of no less than 1 kHz. The physical basis for determining whether the suspended wheel has successfully made contact with the ground is that at the instant the wheel makes contact with the ground from the air, the servo drive torque will increase sharply due to the sudden intervention of the ground reaction force, that is, the time derivative of the torque will show a significant positive step at the moment of contact.

[0163] For example, in this embodiment, the grounding trigger judgment logic of the servo layer can be as follows:

[0164]

[0165] in, for The measured value of the servo drive torque at any given time is obtained by converting the measured motor current using the known motor torque constant. The time derivative of the servo torque remains within a small range during the normal descent phase (when the wheel is moving in the air), but it undergoes a significant positive jump at the instant the wheel contacts the ground. The threshold value for the torque derivative used to determine grounding is preset based on the robot's weight and the rocker arm's structural parameters. This is an indicative function. When the torque derivative exceeds the threshold, it takes the value of 1, triggering a stop-downward command. The suspended rocker arm immediately locks in the current position, completing the bottom-finding process.

[0166] Figure 13 The real-time monitoring curve of the servo torque derivative and the grounding event capture diagram during the bottom-finding process in this embodiment are shown. Figure 13 In the middle (a), the servo drive torque curve is shown. Figure 13 In Figure (b), the torque time derivative curve shows that the torque derivative threshold judgment can capture grounding events with microsecond-level accuracy, which is far superior to the traditional criterion that relies on the absolute value of torque.

[0167] Understandably, accurately capturing grounding events with microsecond-level delays and immediately stopping the downward movement can effectively avoid overshoot—that is, the impact load caused by the rocker arm continuing to move downwards after the wheel has already touched the ground. This impact load may damage the wheel hub structure or cause secondary subsidence on soft ground.

[0168] Sub-case 2: If the physical root cause node is classified as unilateral subsidence, perform creep compaction operation.

[0169] When the physical root cause node is classified as unilateral subsidence, the system instructs the rocker arm mechanism on the subsided side to perform a creep compaction operation. The physical principle of this operation is: by the rocker arm performing a periodic reciprocating lifting and lowering motion with a limited amplitude in the vertical direction, pressure is repeatedly applied to the loose soil under the wheel on the subsided side, causing the soil particles to rearrange and gradually compact, thereby increasing the bearing capacity and friction coefficient of this local area until the wheel can regain sufficient traction.

[0170] In this embodiment, the displacement curve of the creep compaction operation is generated by a preset spatial sine wave generator. The spatial sine wave generator outputs a vertical displacement command for the rocker arm that varies sinusoidally with time, according to preset frequency and amplitude parameters. This displacement command drives the rocker arm on the trapped side to perform periodic up-and-down reciprocating motions with a limited amplitude near its current position. In each compaction cycle, the downward pressing phase of the rocker arm applies compressive force to the soil beneath the wheel, while the upward lifting phase releases pressure, allowing the soil particles to stabilize in their new positions.

[0171] In this embodiment, the termination condition for the creep compaction operation is: continuously monitoring the local soil reaction force at the wheel that is stuck (indirectly obtained through the wheel hub torque sensor). When the measured local soil reaction force exceeds the pre-stored slippage critical threshold (i.e. the minimum ground reaction force value that can ensure the wheel provides sufficient traction without slipping), the compaction is determined to be successful, the creep action is stopped and normal driving is resumed.

[0172] In this embodiment, during the execution of the escaping gait sequence, the state transition matrix continuously evaluates whether the chassis has recovered the expected traction. The specific evaluation method is as follows: compare the difference between the actual displacement speed of the chassis and the theoretical linear speed of the wheels to see if it has fallen below a preset slip threshold. If the difference falls below the threshold and remains below the preset confirmation time, it is determined that the chassis has recovered the expected traction, the execution of the escaping gait sequence is terminated, the system exits the distress and entrapment mode, and normal driving control is restored.

[0173] Step 6: Run a reinforcement learning network module with physical nonlinear constraints in the edge control node to output the desired configuration.

[0174] The reinforcement learning network module refers to the offline training policy matrix loaded in the edge control node. This policy matrix is ​​a set of optimized policy parameters obtained through extensive training iterations in an offline simulation environment using a reinforcement learning algorithm. The input to the policy matrix is ​​a state vector containing current terrain elevation information and real-time identified centroid coordinate information, and the output is the desired velocity command vector for each joint of the rocker arm. .

[0175] Understandably, due to the speed command vector These are numerical values ​​output by a neural network within a continuous action space, and it cannot be guaranteed that every component satisfies the robot's actual physical hardware constraints; for example, the peak torque limit of the motor and the yield stress limit of the link material. If directly... Sending the command to the servo layer for execution could potentially cause motor overload damage or plastic deformation of structural components; therefore, it is essential to... Before entering the servo execution phase, a deterministic hardware constraint filter is used for interception and verification.

[0176] In this embodiment, the execution process of the deterministic hardware-constrained filter is as follows: first, the velocity command vector is calculated. If executed directly, the expected peak motor torque and expected link force values ​​will be calculated based on the known motor torque constant and the link stress distribution model under the current configuration. Then, the expected peak motor torque is compared with the peak torque limits of each motor pre-stored in the hardware register, and the expected link force values ​​are compared with the pre-stored link material yield stress limit. If any expected value exceeds the corresponding hardware limit value, the speed command vector needs to be forcibly corrected.

[0177] The mathematical essence of the forced correction process is a convex quadratic programming problem, that is, finding the distance from the original instruction while satisfying all physical constraints. Recent feasible instructions .

[0178] For example, in this embodiment, the safe and feasible domain is defined as follows:

[0179]

[0180] in, A convex polyhedron is a safe and feasible region defined by physical hardware constraints, and any instruction located inside this region will not cause hardware damage. A linear transformation matrix is ​​used to map joint speed commands to motor torque space. The elements of this matrix are determined by the motor torque constant and transmission ratio of each joint. The peak torque limit vector for each motor is pre-stored in a hardware register; To describe each link in the current configuration The stress distribution matrix under the following stress conditions is updated in real time with the rocker arm's attitude; is the yield stress limit vector of the connecting rod material.

[0181] For example, in this embodiment, the quadratic programming problem of constrained projection can be expressed as follows:

[0182]

[0183] in, The original velocity command vector output by the reinforcement learning network module; The final safety command vector output after constraint projection ensures that it will not violate physical hardware constraints under any operating condition. The square of the weighted norm, It is a weighted matrix used to differentiate the penalties for command deviations in different degrees of freedom—for example, assigning higher weights to the main rocker arm joint that bears critical gravity, thereby prioritizing the protection of critical joints. Let represent the independent variable that minimizes the objective function.

[0184] Understandably, due to the secure feasible domain The given quadratic programming problem is a convex polyhedron defined by a set of linear inequality constraints. It is a rigorous convex optimization problem with a globally unique optimal solution, solvable in real-time within the hardware capabilities of the edge control nodes using efficient effective set methods or interior point methods. Its physical meaning is: to faithfully execute the action policy given by reinforcement learning as much as possible, while ensuring that all robot hardware is not damaged. When it is already within the safe and feasible domain, Constrained projection does not change the original instruction; only when Only when certain components violate hardware constraints is dimensionality reduction truncation performed proportionally to enforce physical boundary safety.

[0185] Step 7: Establish a closed loop for high-frequency execution and contact force feedback in the servo layer.

[0186] After obtaining safe and feasible joint commands through the above steps, the final execution task falls on the hardware control loop within the underlying servo driver of each rocker arm. Unlike edge control node algorithms that run on top of the operating system, the control logic of the servo layer runs on field-programmable gate array hardware, with a control cycle on the order of microseconds, completely bypassing the time uncertainty caused by the task scheduling of the operating system.

[0187] In this embodiment, the servo layer establishes a double-buffered verification mechanism independent of the edge control nodes. The double-buffered verification mechanism works as follows:

[0188] The first buffer is used to receive the expected contact force target value from the edge control node (e.g., determined from the asymmetric compensation command in step S3). The target value is flushed and written to the first buffer during the upper-level control cycle update.

[0189] The second buffer is used to cyclically write the actual contact force sampling values ​​from each rocker arm torque sensor. The sampling value is updated much more frequently than the output frequency of the upper-level controller, ensuring that the servo layer always has the latest force feedback information.

[0190] The FPGA hardware logic gate array inside the servo driver directly compares the difference between the target value in the first buffer and the measured value in the second buffer. Without going through the operating system bus, it hardwires the switching frequency of the motor's three-phase inverter bridge with a microsecond delay to achieve high-frequency force control closed loop.

[0191] For example, in this embodiment, the PWM instruction calculation for the FPGA dual-buffer force control closed loop can be shown as follows:

[0192]

[0193] in, for The pulse width modulation (PWM) duty cycle command vector is constantly output to each servo motor driver. This command directly controls the on / off timing of the power switching transistors in the three-phase inverter bridge of the motor, thereby controlling the output torque of the motor. The proportional gain matrix determines the strength of the force control loop's response to the current force tracking error. The larger the gain, the faster the response, but too large a gain may introduce oscillations. The differential gain matrix determines the damping characteristics of the force control loop on the rate of change of force tracking error. The introduction of the differential term can suppress the oscillations caused by the excessive response of the proportional term and improve the dynamic stability of the system. The target contact force vector, determined by the upper-level compensation model, is stored in the first buffer of the FPGA; The real-time grounding force measurement value is continuously written to the second buffer of the FPGA by the hub sensor at high frequency; This is the force tracking error vector at the current moment. Figure 14 The diagram shows a comparison of the force tracking errors between the FPGA dual-buffered force control closed loop and the operating system-level force control in this embodiment.

[0194] Step 8: Provide adaptive degradation logic for communication in extreme network segmentation scenarios.

[0195] Network segmentation refers to an extreme scenario where the communication link between the edge control node and the cloud processing node is interrupted due to network failures, wireless signal obstruction, or communication base station overload. In this scenario, the edge control node can no longer receive updates of macro-planning instructions from the cloud-based large model, nor can it upload the latest perception data to the cloud to obtain updated experience maps.

[0196] In this embodiment, the communication adaptive degradation logic specifically includes:

[0197] Edge control nodes periodically send heartbeat packets to cloud processing nodes. A heartbeat packet is a lightweight network probe data packet whose sole purpose is to confirm the connectivity of the communication link.

[0198] When heartbeat packets are lost for more than three consecutive cycles, the edge control node determines that the current communication link with the cloud has been interrupted and automatically blocks the macro-planning instruction input interface of the cloud-based large model. The purpose of this blocking operation is to prevent the edge control node from continuing to execute cloud instructions received before the communication interruption, which may be outdated and no longer applicable to the current environment, thus avoiding erroneous interference from outdated instructions on real-time control decisions.

[0199] The system then reverted to local isolated operation mode. In this mode, the system made two key security policy adjustments:

[0200] The first adjustment is: restrict the retrieval scope of the embodied experience memory map in step 2 to the nearest node within the circular buffer residing in the local memory of the edge control node. The trajectory is recorded. The circular buffer is a fixed-size First-In-First-Out (FIFO) data structure that continuously stores the robot's most recently accessed data. The signature vector and corresponding configuration parameters for sub-terrain access. This constraint means that in silo mode, the system relies only on the most recent limited historical experience for terrain matching, rather than the complete cloud experience map, thus avoiding system response delays or anomalies caused by attempts to access unreachable cloud storage.

[0201] The second adjustment is: Adjust the stability margin threshold in step 3. Add a set safety margin constant The threshold will be modified to Increasing the stability margin threshold means that the system will trigger active compensation at a point further away from the support boundary from the center of gravity projection point, thereby sacrificing some driving speed and passability in exchange for a greater rollover safety margin.

[0202] Understandably, the design philosophy behind the aforementioned degradation strategy is to prioritize the robot's physical safety (absolutely preventing tipping over) under extreme conditions where cloud computing support is lost, while allowing for moderate compromises in driving performance (speed and path optimization). This strategy ensures that even in a completely isolated network environment, the robot can still maintain safe and controllable autonomous operation relying solely on local edge and server computing resources. Figure 15 This embodiment shows the impact curves of the stability margin threshold change on the rollover probability in both the isolated offline mode and the normal network mode. Figure 15 As can be seen, in network outage scenarios where edge computing power is limited, actively increasing the stability margin threshold (rather than passively waiting for a response) can effectively reduce the probability of a rollover.

[0203] Example 2:

[0204] This embodiment discloses an adaptive rocker arm handling robot control system based on embodied intelligence. Specifically, this system can be integrated into the robot body and its supporting electronic devices, such as edge computing nodes and cloud servers. The edge computing nodes can be industrial control computers, embedded real-time controllers, or robot-embedded onboard computing platforms. The cloud server can be a single server or a server cluster composed of multiple servers, used to handle computationally intensive tasks such as offline learning, experience graph construction, and macro-planning. When the electronic devices are running, they can implement the adaptive rocker arm handling robot control method based on embodied intelligence as described in Embodiment 1.

[0205] In some embodiments, the adaptive rocker arm handling robot control system based on embodied intelligence can also be integrated into multiple electronic devices. For example, the control system can be deployed according to a three-level distributed architecture of cloud processing node—edge control node—servo driver, with multiple computing units collaboratively implementing the adaptive control method described in this application. The cloud processing node is responsible for offline training, experience map updates, and macro-level task planning; the edge control node is responsible for real-time state estimation, dynamic compensation solving, and safety constraint filtering; and the field-programmable gate array (FPGA) hardware logic inside the servo driver is responsible for the final execution of the microsecond-level high-frequency force control closed loop.

[0206] In some embodiments, the server may also be implemented as an embedded edge computing device to meet the needs of extreme operating scenarios such as on-site deployment and autonomous operation without network access.

[0207] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A control method for an adaptive rocker arm-type transport robot based on embodied intelligence, characterized in that, The control method includes: The first perception dataset and the second state dataset are acquired by a multimodal sensor array mounted on the robot body and rocker arm mechanism. The first perception dataset includes: terrain elevation data and material reflectivity data; the second state dataset includes: servo feedback current of each rocker arm servo motor and current joint angle of each joint. The first perception dataset is converted into a feature vector, and the feature vector is matched with a pre-constructed embodied experience memory map by a distance metric. When a match is found, the historical configuration matrix is ​​extracted and a feedforward configuration instruction is output to achieve feedforward pre-tuning. During obstacle crossing and load loading, based on the normal support force matrix of each rocker arm suspension point and the absolute tilt angle matrix of each rocker arm joint, a system of linear moment balance equations containing unknown load mass and unknown center of gravity coordinates is established and solved to obtain the three-dimensional center of gravity coordinates of the load. Based on the three-dimensional centroid coordinates of the load, a dynamic support polygon is constructed, and the stability margin of the overall centroid projection coordinates is calculated. When the stability margin is lower than the stability margin threshold, an asymmetric compensation command is generated and executed to drive the rocker arms on both sides to perform asymmetric extension and retraction and lifting. During the period of distress and entrapment, a physical directed acyclic graph is constructed based on the causal dependencies between various physical quantities in the robot's motion system. By calculating the normalized residuals of each node and performing topological sorting and traversal along the directed edges, the physical root cause nodes leading to the loss of power are isolated. Based on the attributes of the physical root cause nodes, the corresponding mimicry gait sequence for escaping difficulties is matched from the state transition matrix and the escaping action is executed. A reinforcement learning network module is run in the edge control node to output the original speed command vector. The original speed command vector is projected into the safe and feasible region defined by the motor torque limit and the connecting rod yield stress limit to obtain the safe command vector and issue it for execution.

2. The adaptive rocker arm handling robot control method based on embodied intelligence according to claim 1, characterized in that, The step of converting the first perceptual dataset into a feature vector specifically includes: The spatial area covered by the aforementioned terrain elevation data is uniformly divided into multiple three-dimensional voxel grids of fixed size; Local plane fitting is performed on the point cloud sampling points contained in each three-dimensional voxel mesh to estimate the surface normal vector at each sampling point, and the variance of all surface normal vectors in the three-dimensional voxel mesh is calculated to obtain the surface normal vector variance of the three-dimensional voxel mesh. Extract the corresponding reflection intensity values ​​of the point cloud sampling points contained in each three-dimensional voxel mesh and calculate the arithmetic mean to obtain the average material reflectivity of the three-dimensional voxel mesh. The surface normal variance and the mean material reflectance of each of the three-dimensional voxel meshes are concatenated and stitched together in voxel order to generate a fixed-dimensional terrain signature vector, which is used as the feature vector. The distance metric matching includes: In the edge control node, the terrain signature vector is used as a query key and input into the embodied experience memory graph constructed based on the locality-sensitive hashing algorithm for retrieval; wherein, The embodied experience memory map contains multiple independent random projection hash functions. Each random projection hash function maps the terrain signature vector to an integer bucket number by performing an inner product operation between the terrain signature vector and a random projection vector, adding a random offset, dividing by the quantization slot width parameter and rounding. When a match is found, the process of extracting the historical configuration matrix and outputting feedforward configuration instructions includes: The proportion of the terrain signature vector and the historical terrain signature vector stored in the embodied experience memory map that have the same bucket number on all random projective hash functions is calculated, and the complement of this proportion is used as the hash dissimilarity. When the hash dissimilarity is lower than the first threshold, a match is determined, the historical configuration matrix is ​​extracted from the corresponding hash bucket, and the historical configuration matrix is ​​output as the feedforward configuration instruction. When the hash dissimilarity is not lower than the first threshold, the real-time solution process is initiated.

3. The adaptive rocker arm handling robot control method based on embodied intelligence according to claim 1, characterized in that, The three-dimensional centroid coordinates of the load are obtained through the following steps: Read the normal support force matrix obtained by force sensors at each wheel contact point or by mapping the servo feedback current through the motor torque constant, and the absolute tilt angle matrix fed back by the absolute position encoder built into each rocker arm joint; Based on the absolute tilt angle matrix, the three-dimensional spatial coordinates of each wheel contact point are obtained through forward kinematics calculations. A spatial moment balance equation is then established, incorporating the unknown load mass and the three-dimensional centroid coordinates of the unknown load. The sum of the torques generated by the normal support forces at each wheel contact point is equal to the sum of the torques generated by the robot's known weight and the unknown load under the action of gravity. The product of the unknown load mass and the three-dimensional centroid coordinates of the unknown load in the spatial moment balance equation is defined as a component of the load state vector. The spatial moment balance equation is rewritten as a system of linear moment balance equations with respect to the load state vector; The singular value decomposition algorithm is used to solve the linear equations of torque balance to obtain the optimal load state vector; The load mass component is extracted from the optimal load state vector, and the remaining components are divided by the load mass component to obtain the three-dimensional centroid coordinates of the load.

4. The adaptive rocker arm handling robot control method based on embodied intelligence according to claim 3, characterized in that, The optimal load state vector is obtained through the following steps: Perform singular value decomposition on the coefficient matrix of the linear equation system of torque balance to obtain a left orthogonal singular vector matrix, a singular value diagonal matrix, and a right orthogonal singular vector matrix. Take the reciprocal of the singular values ​​in the singular value diagonal matrix whose values ​​are greater than a preset truncation threshold, and set the singular values ​​whose values ​​are not greater than the preset truncation threshold to zero to obtain the truncated generalized inverse matrix. The optimal load state vector is obtained by multiplying the right-hand orthogonal singular vector matrix, the truncated generalized inverse matrix, and the transpose of the left-hand orthogonal singular vector matrix with the right-hand vector of the torque balance linear equation system.

5. The adaptive rocker arm handling robot control method based on embodied intelligence according to claim 1, characterized in that, The generation and execution of the asymmetric compensation command includes: Take the horizontal component of the three-dimensional coordinates of all bearing wheel ground points, and obtain the convex hull of the planar point set formed by the horizontal component to obtain the dynamic support polygon; Based on the known coordinates of the robot's own center of gravity and the three-dimensional center of gravity of the load, the vertical projection coordinates of the overall system's center of gravity on the horizontal plane are calculated to obtain the overall center of gravity projection coordinates. Calculate the normal distance from the overall centroid projection coordinates to the boundary lines of each boundary line of the dynamic support polygon, and take the minimum value as the stability margin. When the stability margin is lower than the stability margin threshold, the nearest boundary corresponding to the stability margin is determined as the potential overturning axis, and a horizontal displacement vector of the center of gravity is generated from the overall center of gravity projection coordinates to the geometric center of the dynamic support polygon. By using the pseudo-inverse of the Jacobian matrix of the rocker arm system, the horizontal displacement vector of the center of gravity is mapped to the joint space to obtain the asymmetric elongation and lifting angle adjustment increment of each rocker arm, which are then issued and executed as the asymmetric compensation command.

6. The adaptive rocker arm handling robot control method based on embodied intelligence according to claim 1, characterized in that, The physical root causes of the isolation leading to a lack of power include: Define the set of nodes in the physical directed acyclic graph, where each node represents an observable physical quantity, and the set of nodes includes at least the chassis acceleration node, the wheel speed nodes of each wheel, and the servo current nodes of each rocker arm. Define the set of directed edges of the physical directed acyclic graph, where each directed edge represents a physical causal dependency from cause to effect; For each node in the physical directed acyclic graph, the theoretical expected value of the physical quantity represented by the node is calculated according to the dynamic model, and the normalized residual between the actual measured value of the node and the theoretical expected value is calculated. The nodes are traversed along the directed edges in a topological sort. When the normalized residual of a node exceeds the fault determination threshold, and the normalized residual of all direct parent nodes of the node in the physical directed acyclic graph does not exceed the normal range determination threshold, the node is determined to be the physical root cause node.

7. The adaptive rocker arm handling robot control method based on embodied intelligence according to claim 6, characterized in that, The fault types of the physical root cause nodes are classified into one of the following: Diagonal suspension, unilateral sinking, or chassis bottoming; The step of matching the corresponding mimicry gait sequence for escaping difficulty from the state transition matrix based on the attributes of the physical root cause nodes and executing the escaping action includes: Based on the fault type classification results of the physical root cause nodes, key-value matching is performed in the state transition matrix to obtain the corresponding simulated escape gait sequence parameters and execute them; When the fault type is "diagonal line suspended", perform a bottom-seeking operation: Lock the brakes on the wheels that are still in contact with the ground. The command arm mechanism on the suspended side continuously increases its downward stroke at a preset slow speed. During the descent, the servo layer continuously monitors the derivative of the contact torque of the suspended servo joint at a sampling frequency no less than the preset frequency. When the derivative of the contact torque exceeds the grounding determination threshold, it is determined that the suspended wheel has contacted the ground, the downward movement is immediately stopped and the rocker arm on that side is locked in the current position. When the fault type is unilateral subsidence, perform creep compaction: The rocker arm mechanism on the command side performs periodic reciprocating lifting and lowering movements in the vertical direction according to a preset frequency and preset amplitude. During the downward pressing phase of each compaction cycle, it applies compressive force to the soil under the wheel and releases pressure during the upward lifting phase. The local soil reaction force at the side wheel is continuously monitored. When the local soil reaction force exceeds the pre-stored slippage critical threshold, the creep compaction operation is stopped and normal drive is resumed.

8. The adaptive rocker arm handling robot control method based on embodied intelligence according to claim 1, characterized in that, The security instruction vector is obtained through the following steps: Based on the known motor torque constant and the linkage stress distribution model under the current configuration, calculate the expected peak motor torque and expected linkage force value when the original speed command vector is directly executed; The expected peak motor torque is compared with the pre-stored peak torque limits of each motor, and the expected connecting rod force value is compared with the pre-stored connecting rod material yield stress limit. If any expected value exceeds the corresponding limit, then using convex quadratic programming, within the safe and feasible region jointly defined by the motor torque limit constraint and the connecting rod yield stress limit constraint, the feasible command with the smallest weighted norm distance to the original speed command vector is selected as the safe command vector.

9. The adaptive rocker arm handling robot control method based on embodied intelligence according to claim 1, characterized in that, The control method further includes: At the servo layer, a double-buffered verification mechanism independent of the edge control nodes is established, wherein... The first buffer receives the target contact force from the edge control node, and the second buffer is cyclically written with the real-time contact force sampling values ​​of each rocker arm torque sensor. The servo driver's internal field-programmable gate array (FPGA) hardware logic directly compares the difference between the target contact force in the first buffer and the real-time contact force sample value in the second buffer. Based on the difference and its time rate of change, it generates a pulse width modulation (PWM) duty cycle command using proportional-derivative (PD) control to drive each servo motor to execute a force control closed loop. It also includes communication adaptive degradation logic: The edge control node periodically sends heartbeat packets to the cloud processing node; When the heartbeat packets are lost for more than a preset number of cycles, the edge control node determines that the communication link with the cloud processing node is interrupted, automatically blocks the instruction input interface from the cloud processing node, and switches to local island operation mode. In the local isolated operation mode, the retrieval range of the embodied experience memory map is limited to the most recent preset number of trajectory records in the circular buffer residing in the local memory of the edge control node, and the stability margin threshold is increased by a safety margin constant.

10. A control system for an adaptive rocker arm-type transport robot based on embodied intelligence, characterized in that, The control system includes: processor; The memory stores a computer program that, when executed by a processor, implements the adaptive rocker arm handling robot control method based on embodied intelligence as described in any one of claims 1 to 9.

Images