Wheel-legged robot motion mode smooth switching control method

Abstract

The application discloses a kind of based on multimodal perception and weighted null space mapping's smooth mode switching control method of wheel-legged robot, it is related to robot control technical field.For the problem that existing wheel-legged robot is difficult to autonomously smooth switching and motion mode switching is not coherent under complex terrain, the application fuses IMU, laser radar and camera data, and utilizes vision language model (VLM) to estimate the terrain related parameters of contact surface comprehensively, constructs terrain feature vector.Based on the vector, mapping function is designed to generate joint weight matrix to represent the contribution of each joint in different terrain tasks.Furthermore, the weight matrix is introduced into the null space control law to construct the weighted pseudo-inverse, and the expected motion of each joint is solved and continuously differentiable processing is carried out.The application effectively eliminates the mechanical impact during mode switching, and realizes the autonomous, smooth and stable switching of wheel-legged form in complex unknown environment such as post-disaster.

Description

Technical Field

[0001] This invention relates to the field of wheeled robot control, and specifically to a method for smooth switching of motion modes in wheeled robots based on multimodal perception and weighted null space mapping. Background Technology

[0002] Wheeled-legged robots combine the high speed and energy efficiency of traditional wheeled robots with the terrain adaptability and obstacle-crossing ability of legged robots. In extremely complex and obstacle-filled unknown environments such as disaster relief and field exploration, wheeled-legged robots have demonstrated enormous application potential. However, existing wheeled-legged robots still suffer from the following significant shortcomings in their underlying control strategies for autonomous environmental perception and motion mode switching:

[0003] Existing terrain perception systems primarily rely on pure vision or pure LiDAR for geometric feature modeling. Geometric perception is prone to failure under extreme conditions such as dust storms or sudden changes in lighting. Furthermore, traditional methods struggle to obtain accurate physical semantic parameters such as local ground roughness and friction coefficients, leading to a lack of anticipation of complex and variable road surfaces like "soft swamps" and "dark ice," making the robot highly susceptible to wheel slippage or foot sinking. When faced with complex terrain triggering wheel-foot mode transitions, existing control strategies often employ state machine-based "hard switching." This approach typically requires the robot to pause or execute a pre-set rigid gait transition, causing a sudden abrupt change in the underlying control inputs—joint acceleration and driving torque—at the moment of transition. This discontinuous dynamic output not only accelerates wear on mechanical structures but also causes the robot to lose balance on slopes or rough surfaces. In traditional whole-body motion control based on zero-space technology, the weight allocation matrix for each joint when performing the main task and sub-tasks is usually fixed or adjusted offline based on experience. When the robot faces a dynamic change in the road environment, such as suddenly moving from a high-friction, smooth cement road to a low-friction, muddy grass, the system cannot adaptively adjust the output contribution of the wheel drive joints and the leg motion joints online, which limits the terrain adaptation potential and collaborative efficiency of the wheel-legged robot. Summary of the Invention

[0004] The technical solution adopted in this invention is a smooth switching control method for the motion modes of a wheeled robot based on multimodal perception and weighted null-space mapping. Based on data collected from multiple sources of sensors and a visual language model, the terrain feature vector of the road surface is calculated. The weight matrix of each joint of the robot is obtained by mapping the terrain feature vector. This weight matrix is ​​then introduced into the pseudo-inverse of the robot's null-space control law. The weighted pseudo-inverse of the weight matrix is ​​used to solve for the acceleration of different joints during task execution, thus achieving autonomous and smooth control of the wheeled robot's motion modes under different road conditions. Specifically, the method includes the following steps:

[0005] Step 1: Acquire multi-source sensor data of the wheeled robot's operating environment, including inertial measurement unit (IMU) data, lidar point cloud data, and visual image data captured by the camera;

[0006] Step 2: The multi-source sensor data is fused to calculate the geometric features such as the slope of the road surface, and the visual language model (VLM) is used to identify terrain features such as grass, cement road or swamp. Based on the identified terrain, the physical parameters of the robot's contact surface are estimated, and a terrain feature vector is constructed.

[0007] Step 3: Design a weight mapping function to map the terrain feature vector into a joint weight matrix. The joint weight matrix represents the contribution of different joints of the robot in the execution of the current terrain task.

[0008] Step 4: Introduce the joint weight matrix into the robot's null space control law, construct a null space control law based on weighted pseudo-inverse, define the robot's overall motion as the main task, and use the weighted pseudo-inverse introduced by the weight matrix to solve for the expected velocity and expected acceleration of each joint.

[0009] Step 5: Perform continuous differentiable processing on the control input to generate the final motor drive torque, and control the robot to perform smooth switching between wheeled, legged, or hybrid wheeled and legged modes.

[0010] Furthermore, in step two, the specific method for constructing the terrain feature vector includes:

[0011] Terrain feature vector Including local slope obstacle height Terrain roughness and The calculation methods for each parameter are as follows:

[0012] Based on the linear acceleration and angular velocity data of the robot body coordinate system collected by the IMU, the motion acceleration interference is filtered out and the gravitational acceleration component is separated by the attitude calculation algorithm. Combined with the robot kinematic model, coordinate system transformation is performed to calculate the robot's current attitude. The original 3D point cloud of the LiDAR is then pose-compensated and uniformly transformed to the world coordinate system.

[0013] Based on the robot's current orientation, a local point cloud set in front of the robot is dynamically extracted from the global point cloud map scanned by LiDAR. .

[0014] Calculate the centroid of a local point cloud And construct the covariance matrix :

[0015]

[0016] right Perform singular value decomposition; the eigenvector corresponding to the smallest eigenvalue is the normal vector of the fitted plane. Then calculate the local slope. :

[0017]

[0018] obstacle height This characterizes the geometric span of the plane within a local area. A local point cloud within a preset range in front of the robot is extracted, and the local support plane is fitted using the RANSAC algorithm as a reference. Then, the orthogonal distance from each sampling point in the point cloud set to the reference is calculated, and the difference between the maximum and minimum orthogonal distances within this region is defined as the real-time obstacle height. .

[0019]

[0020] roughness It represents the degree of undulation of the ground. Calculate the point cloud set. The perpendicular distance of all points to the above-mentioned fitting plane Roughness is defined as the standard deviation of these distances:

[0021]

[0022] Visual image data captured by the camera is input into the visual language model, which outputs terrain category and probability. A hash mapping table of terrain semantics and physical parameters is pre-built in memory. The keys of the hash table are defined as terrain semantic strings output by the VLM, such as "grass", "concrete", "swamp", "land", etc., and the values ​​are defined as values ​​containing the friction coefficient. After the visual language model outputs the terrain category, the friction coefficient corresponding to the terrain can be obtained from a table immediately. .

[0023] The specific values ​​in the mapping table are not randomly assigned, but obtained through offline experimental calibration. This allows the wheeled robot to undergo dragging or slipping tests on typical road surfaces, recording the corresponding maximum static friction and vertical vibration acceleration, thereby calculating the baseline values. Store in memory.

[0024] After the multi-source sensors acquire data, the terrain feature vector can be obtained in real time based on the above calculation process.

[0025] Furthermore, the method for mapping the terrain feature vector to a joint weight matrix in step three includes:

[0026] First, for each parameter in the terrain feature vector, a "wheel-type attenuation / foot-type excitation" function between 0 and 1 needs to be designed. These functions must be smooth and continuously differentiable. The evaluation functions for the four parameters of the feature vector are designed as follows:

[0027] The lower the coefficient of friction (e.g., on ice or mud), the easier it is for the wheels to slip, resulting in lower wheel efficiency. Therefore, linear or smooth saturated mapping is used.

[0028]

[0029] An increased gradient causes a sinusoidal increase in the driving torque required by the wheeled vehicle, which can easily lead to slippage or overturning. Therefore, a cosine attenuation model is used:

[0030]

[0031] Roughness causes continuous high-frequency vibrations, consuming wheel energy. An exponential decay model is adopted, where... The roughness sensitivity coefficient controls the degree of attenuation.

[0032]

[0033] When the height of the obstacle is close to the radius of the wheel At this time, the wheel absolutely cannot pass. A forced interception must be performed using a Sigmoid function with soft boundaries.

[0034]

[0035] By integrating these four univariate evaluation factors, the overall confidence level or contribution of the current environment to wheeled sports can be calculated. :

[0036]

[0037] Calculate the overall confidence level or contribution of foot movements. for:

[0038]

[0039] Weighted null space control In the mathematical definition, the weight matrix Located in the denominator or inverse term of the quadratic optimization, meaning the larger the weight value, the greater the penalty for the joint's movement, and the more prone the joint is to locking. An improved Sigmoid function is used to map the contribution to penalty weights in the joint augmented Jacobian matrix:

[0040] The penalty weights for wheel-type and foot-type joints can be obtained as follows:

[0041]

[0042] Where w_max is the maximum penalty limit set, with a value between 50 and 100, and λ is the switching sensitivity gain. Assuming a single leg has three joints (hip, knee, and ankle) and one wheel at the end, the joint weight matrix for that leg is constructed as follows:

[0043]

[0044] Furthermore, the specific process of constructing the control law based on the weighted pseudo-inverse in step four is as follows:

[0045] Assume the robot has one leg with joints including hip lateral movement, hip pitch, and knee joint, and an end effector connected to a drive wheel. Define the leg joint variables as follows: Define the rolling joint variable of the wheel as Therefore, the complete generalized joint coordinates of this single-leg-wheel system are:

[0046]

[0047] Jacobian matrix Describes joint velocity With end effector linear velocity Linear mapping relationship between them:

[0048]

[0049] Within this framework, the velocity at the point of contact is the superposition of two components: the translation caused by leg joint movement and the rolling translation caused by wheel rotation. Leg kinematic components. Considering only the linear velocity contribution of each leg joint motion to the wheel center, this can be obtained by taking the partial derivative of the robot's forward kinematics with respect to the leg variables. Wheel kinematic components. Let the radius of the wheel be... The unit vector of the tangent direction in the plane containing the wheel is The contribution of wheel rolling to the linear velocity at the contact point is... Therefore, the Jacobian column vector of the wheel is... By concatenating the two, we obtain the unified Jacobian matrix:

[0050]

[0051] Introducing the previously constructed diagonal weight matrix The formula for calculating the weighted pseudoinverse is:

[0052]

[0053] Based on the weighted pseudo-inverse matrix, an improved velocity-level null-space control law is constructed, and the calculation formula is as follows:

[0054]

[0055] in, For the desired joint velocity, For Jacobian matrices, For weighted pseudo-reverse, The expected speed of the main task, The weighted null space projection matrix ensures that the execution of subtasks is completely mapped within the null space of the main task, meaning that the movement of subtasks will not interfere with the main task. The expected joint velocity for the subtask. It is an identity matrix.

[0056] To further eliminate sudden torque changes during mode switching and ensure a smooth response from the underlying motor controller, the aforementioned velocity-level control law is differentiated to construct an acceleration-level weighted zero-space control law:

[0057]

[0058] in, The main task expects acceleration. The time derivative of the Jacobian matrix. The desired acceleration for the subtask. The system can express the desired acceleration using the inverse dynamic equations. This is converted into the desired torque for each joint and sent to each joint actuator.

[0059] Furthermore, the method for performing continuously differentiable processing of the control input in step five is as follows:

[0060] To ensure the continuous differentiability of the control input (i.e., the continuity of the first derivative) and avoid abrupt acceleration changes, a cubic spline curve is used as the transition smoothing function. A normalized time variable is defined. ,in Construct a smooth transition factor:

[0061]

[0062] The function satisfies And satisfy During the transition time window Internally, the continuous weight matrix actually distributed to the zero-space control law :

[0063]

[0064] When the time exceeds the transition window back, At this point, the system transitions to the new terrain very smoothly.

[0065] The above time-domain filtered Substituting into the weighted null space control law in step four, the continuous and abruptly-free desired joint velocity is obtained. and desired joint acceleration .Will and Introducing the rigid body dynamics equations of a wheeled robot, we perform inverse dynamics feedforward calculations. The standard dynamics equations of the robot are expressed as:

[0066]

[0067] in The final desired motor drive torque vector is calculated as follows: Let be the symmetric positive definite inertia matrix of the robot. The matrix of Coriolis force and centrifugal force. Let the gravity compensation term vector be... This is a term for compensation of joint friction.

[0068] Due to the input At least have Continuity, calculated acceleration It is continuously distributed. Therefore, the final torque command derived through the mapping of the above dynamic equations... It is also continuous and smooth.

[0069] This invention provides a control method for smooth switching of wheel-foot modes in a wheeled robot based on weighted null space mapping. Compared with existing technologies, the advantages of this invention are as follows:

[0070] 1. This invention introduces a visual language model (VLM) for terrain recognition reasoning. Combined with IMU and LiDAR, it can not only identify the geometric protrusions of the terrain, but also directly calculate the local roughness and friction coefficient.

[0071] 2. This invention establishes a mathematical model that maps terrain feature vectors to joint weight matrices. When the terrain changes, the system can adjust the motion contribution of wheeled and legged joints, breaking the discrete adjustment mode of traditional control and fully leveraging the terrain adaptability potential of wheeled and legged robots.

[0072] 3. This invention introduces a dynamic weight matrix into pseudo-inverse calculation and constructs a weighted zero-space control law, which seamlessly completes the leg movements and wheel takeover using redundant degrees of freedom without interrupting the robot's main task. Attached Figure Description

[0073] Figure 1 This is a flowchart illustrating the implementation of the method of the present invention.

[0074] Figure 2 To simulate the changes in the parameters of the terrain feature vector when the terrain changes.

[0075] Figure 3 This describes the parameter changes during the mapping from terrain feature vectors to weight matrices.

[0076] Figure 4 This describes the numerical change of the norm of the weighted pseudo-inverse matrix during mode switching.

[0077] Figure 5 This describes the process of the output torque value change of the underlying drive motor. Detailed Implementation

[0078] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0079] A method for smooth switching of motion modes in a wheeled robot based on multimodal perception and weighted null-space mapping is proposed. This method calculates terrain feature vectors from data collected by multiple sensors and a visual language model. Weight matrices for each joint of the robot are obtained by mapping these terrain feature vectors. These weight matrices are then incorporated into the pseudo-inverse of the robot's null-space control law. The weighted pseudo-inverse of the weighted matrix is ​​used to solve for the acceleration of different joints during task execution, thus achieving autonomous and smooth control of the wheeled robot's motion modes under different road conditions. The specific steps include:

[0080] Step 1: Acquire multi-source sensor data of the wheeled robot's operating environment, including inertial measurement unit (IMU) data, lidar point cloud data, and visual image data captured by the camera;

[0081] Step 2: The multi-source sensor data is fused to calculate the geometric features such as the slope of the road surface, and the visual language model (VLM) is used to identify terrain features such as grass, cement road or swamp. Based on the identified terrain, the physical parameters of the robot's contact surface are estimated, and a terrain feature vector is constructed.

[0082] Step 3: Design a weight mapping function to map the terrain feature vector into a joint weight matrix. The joint weight matrix represents the contribution of different joints of the robot in the execution of the current terrain task.

[0083] Step 4: Introduce the joint weight matrix into the robot's null space control law, construct a null space control law based on weighted pseudo-inverse, define the robot's overall motion as the main task, and use the weighted pseudo-inverse introduced by the weight matrix to solve for the expected velocity and expected acceleration of each joint.

[0084] Step 5: Perform continuous differentiable processing on the control input to generate the final motor drive torque, and control the robot to perform smooth switching between wheeled, legged, or hybrid wheeled and legged modes.

[0085] In step two, the specific method for constructing the terrain feature vector includes:

[0086] Terrain feature vector Including local slope obstacle height Terrain roughness and The calculation methods for each parameter are as follows:

[0087] Based on the linear acceleration and angular velocity data collected by the IMU in the robot's body coordinate system, the motion acceleration interference is filtered out and the gravitational acceleration component is separated through the attitude calculation algorithm. Combined with the robot's kinematic model, coordinate system transformation is performed to calculate the robot's current attitude. The original 3D point cloud of the LiDAR is then pose-compensated and uniformly transformed to the world coordinate system.

[0088] Based on the robot's current orientation, a local point cloud set in front of the robot is dynamically extracted from the global point cloud map scanned by LiDAR. .

[0089] Calculate the centroid of a local point cloud And construct the covariance matrix :

[0090]

[0091] right Perform singular value decomposition; the eigenvector corresponding to the smallest eigenvalue is the normal vector of the fitted plane. Then calculate the local slope. :

[0092]

[0093] obstacle height This characterizes the geometric span of the plane within a local area. A local point cloud within a preset range in front of the robot is extracted, and the local support plane is fitted using the RANSAC algorithm as a reference. Then, the orthogonal distance from each sampling point in the point cloud set to the reference is calculated, and the difference between the maximum and minimum orthogonal distances within this region is defined as the real-time obstacle height. .

[0094]

[0095] Roughness characterizes the degree of undulation in the ground. (Calculate the point cloud set.) The perpendicular distance of all points to the above-mentioned fitting plane Roughness is defined as the standard deviation of these distances:

[0096]

[0097] Visual image data captured by the camera is input into the visual language model, which outputs terrain category and probability. A hash mapping table of terrain semantics and physical parameters is pre-built in memory. The keys of the hash table are defined as terrain semantic strings output by the VLM, such as "grass", "concrete", "swamp", "land", etc., and the values ​​are defined as values ​​containing the friction coefficient. After the visual language model outputs the terrain category, the friction coefficient corresponding to the terrain can be obtained from a table immediately. .

[0098] The specific values ​​in the mapping table are not randomly assigned, but obtained through offline experimental calibration. This allows the wheeled robot to undergo dragging or slipping tests on typical road surfaces, recording the corresponding maximum static friction and vertical vibration acceleration, thereby calculating the baseline values. Store in memory.

[0099] After the multi-source sensors acquire data, the terrain feature vector can be obtained in real time based on the above calculation process.

[0100] The method for mapping terrain feature vectors to joint weight matrices in step three includes:

[0101] First, for each parameter in the terrain feature vector, a "wheel-type attenuation / foot-type excitation" function between 0 and 1 needs to be designed. These functions must be smooth and continuously differentiable. The evaluation functions for the four parameters of the feature vector are designed as follows:

[0102] The lower the coefficient of friction (e.g., on ice or mud), the easier it is for the wheels to slip, resulting in lower wheel efficiency. Therefore, linear or smooth saturated mapping is used.

[0103]

[0104] An increased gradient causes a sinusoidal increase in the driving torque required by the wheeled vehicle, which can easily lead to slippage or overturning. Therefore, a cosine attenuation model is used:

[0105]

[0106] Roughness causes continuous high-frequency vibrations, consuming wheel energy. An exponential decay model is adopted, where... The roughness sensitivity coefficient controls the degree of attenuation.

[0107]

[0108] When the height of the obstacle is close to the radius of the wheel At this time, the wheel absolutely cannot pass. A forced interception must be performed using a Sigmoid function with soft boundaries.

[0109]

[0110] By integrating these four univariate evaluation factors, the overall confidence level or contribution of the current environment to wheeled sports can be calculated. :

[0111]

[0112] Calculate the overall confidence level or contribution of foot movements. for:

[0113]

[0114] Weighted null space control In the mathematical definition, the weight matrix Located in the denominator or inverse term of the quadratic optimization, meaning the larger the weight value, the greater the penalty for the joint's movement, and the more prone the joint is to locking. An improved Sigmoid function is used to map the contribution to penalty weights in the joint augmented Jacobian matrix:

[0115] The penalty weights for wheel-type and foot-type joints can be obtained as follows:

[0116]

[0117] in, The maximum penalty limit is set, with a value between 50 and 100. To switch sensitivity gain.

[0118] Assuming a single leg has three joints (hip, knee, and ankle) and one wheel at its end, the joint weight matrix for that leg is constructed as follows:

[0119]

[0120] The specific process of constructing the control law based on the weighted pseudo-inverse in step four is as follows:

[0121] Assume the robot has one leg with joints including hip lateral movement, hip pitch, and knee joint, and an end effector connected to a drive wheel. Define the leg joint variables as follows: Define the rolling joint variable of the wheel as Therefore, the complete generalized joint coordinates of this single-leg-wheel system are:

[0122]

[0123] Jacobian matrix Describes joint velocity With end effector linear velocity Linear mapping relationship between them:

[0124]

[0125] Within this framework, the velocity at the point of contact is the superposition of two components: the translation caused by leg joint movement and the rolling translation caused by wheel rotation. Leg kinematic components. Considering only the linear velocity contribution of each leg joint motion to the wheel center, this can be obtained by taking the partial derivative of the robot's forward kinematics with respect to the leg variables. Wheel kinematic components. Let the radius of the wheel be... The unit vector of the tangent direction in the plane containing the wheel is The contribution of wheel rolling to the linear velocity at the contact point is... Therefore, the Jacobian column vector of the wheel is... By concatenating the two, we obtain the unified Jacobian matrix:

[0126]

[0127] Introducing the previously constructed diagonal weight matrix The formula for calculating the weighted pseudoinverse is:

[0128]

[0129] Based on the weighted pseudo-inverse matrix, an improved velocity-level null-space control law is constructed, and the calculation formula is as follows:

[0130]

[0131] in, For the desired joint velocity, For Jacobian matrices, For weighted pseudo-reverse, The expected speed of the main task, The weighted null space projection matrix ensures that the execution of subtasks is completely mapped within the null space of the main task, meaning that the movement of subtasks will not interfere with the main task. The expected joint velocity for the subtask. It is an identity matrix.

[0132] To further eliminate sudden torque changes during mode switching and ensure a smooth response from the underlying motor controller, the aforementioned velocity-level control law is differentiated to construct an acceleration-level weighted zero-space control law:

[0133]

[0134] in, The main task expects acceleration. The time derivative of the Jacobian matrix. The desired acceleration for the subtask. The system can express the desired acceleration using the inverse dynamic equations. This is converted into the desired torque for each joint and sent to each joint actuator.

[0135] The method for continuously differentiable processing of the control input in step five is as follows:

[0136] To ensure the continuous differentiability of the control input (i.e., the continuity of the first derivative) and avoid abrupt acceleration changes, a cubic spline curve is used as the transition smoothing function. A normalized time variable is defined. ,in Construct a smooth transition factor:

[0137]

[0138] The function satisfies And satisfy During the transition time window Internally, the continuous weight matrix actually distributed to the zero-space control law :

[0139]

[0140] When the time exceeds the transition window back, At this point, the system transitions to the new terrain very smoothly.

[0141] The above time-domain filtered Substituting into the weighted null space control law in step four, the continuous and abruptly-free desired joint velocity is obtained. and desired joint acceleration .Will and Introducing the rigid body dynamics equations of a wheeled robot, we perform inverse dynamics feedforward calculations. The standard dynamics equations of the robot are expressed as:

[0142]

[0143] in The final desired motor drive torque vector is calculated as follows: Let be the symmetric positive definite inertia matrix of the robot. The matrix of Coriolis force and centrifugal force. Let the gravity compensation term vector be... This is a term for compensation of joint friction.

[0144] Due to the input At least have Continuity, calculated acceleration It is continuously distributed. Therefore, the final torque command derived through the mapping of the above dynamic equations... It is also continuous and smooth.

[0145] To verify the effectiveness of the smooth switching control method for motion modes of a wheeled legged robot based on multimodal perception and weighted null space mapping described in this invention, a dynamics simulation platform based on Python was built. Simulation settings were configured. The step size and total duration are 2.5 seconds. At time 1.2 seconds, the simulated robot moves from a smooth concrete surface to a complex low-friction swamp surface.

[0146] See attached document Figure 2 This demonstrates the terrain feature vectors in step two. The real-time extraction results. In the 0-1.2s segment, the slope obtained by fitting simulated cement site clouds. Stable at Nearby, obstacle height and ruggedness Maintained at an extremely low level. After entering swampy terrain in 1.2 seconds, the ruggedness... Height of the obstacle The speed increased rapidly, and the Visual Language Model (VLM) detected a change in terrain semantics from "concrete" to "swamp" in 1.2 seconds, corresponding to the friction coefficient. The simulation results show that the present invention, by combining spatial geometric features with VLM semantic physical features, can accurately and in real time capture environmental changes, providing a reliable basis for subsequent weight allocation.

[0147] See attached document Figure 3 This demonstrates the mapping process between the evaluation factors and the weight matrix in step three. As the terrain features change, the four evaluation factors... Different trends of change.

[0148] During the transition window (1.2s-1.8s) following the decision trigger, the wheel contribution rate... Contribution of footwork It exhibits a smooth crossover trend.

[0149] See attached document Figure 4 With appendix Figure 5 This verifies the stability of the control law. Figure 4 Display the norm of the weighted pseudoinverse matrix. The invention exhibits a continuous and smooth transition during mode switching, without numerical explosions or drastic jumps, demonstrating its computational robustness in handling redundant degrees of freedom allocation. (See attached...) Figure 5 The output torque value of the underlying drive motor is shown. Thanks to the cubic spline smoothing factor and time-domain filtering processing described in step five, the leg joint torque... With drive wheel torque The switching process is extremely smooth, with continuous first derivatives, eliminating the dynamic shock caused by mode switching.

Claims

1. A method for smooth switching control of motion modes in a wheeled legged robot, characterized in that: Based on data collected from multiple sensors and a visual language model, terrain feature vectors of the road surface are calculated. Weight matrices for each joint of the robot are then obtained by mapping these terrain feature vectors. These weight matrices are introduced into the pseudo-inverse of the robot's null-space control law. The weighted pseudo-inverse of the weighted matrix is ​​then used to solve for the acceleration of different joints during task execution, achieving autonomous and smooth control of the wheeled robot's wheeled motion modes under different road conditions. Specifically, the following steps are included: Step 1: Acquire multi-source sensor data of the wheeled robot's operating environment, including inertial measurement unit (IMU) data, lidar point cloud data, and visual image data captured by the camera; Step 2: The multi-source sensor data is fused to calculate the geometric features such as the slope of the road surface, and the visual language model (VLM) is used to identify the terrain features of grass, cement road or swamp. Based on the identified terrain, the physical parameters of the robot's contact surface are estimated and a terrain feature vector is constructed. Step 3: Design a weight mapping function to map the terrain feature vector into a joint weight matrix. The joint weight matrix represents the contribution of different joints of the robot in the execution of the current terrain task. Step 4: Introduce the joint weight matrix into the robot's null space control law, construct a null space control law based on weighted pseudo-inverse, define the robot's overall motion as the main task, and use the weighted pseudo-inverse introduced by the weight matrix to solve for the expected velocity and expected acceleration of each joint. Step 5: Perform continuous differentiable processing on the control input to generate the final motor drive torque, and control the robot to perform smooth switching between wheeled, legged, or hybrid wheeled and legged modes.

2. The wheel-legged robot motion mode smooth switching control method according to claim 1, characterized in that, In step two, the specific method for constructing the terrain feature vector includes: Terrain feature vector Including local slope obstacle height Terrain roughness and , The transpose sign is used; the parameters are calculated as follows: Based on the linear acceleration and angular velocity data of the robot body coordinate system collected by the IMU, the motion acceleration interference is filtered out and the gravitational acceleration component is separated through the attitude calculation algorithm. Combined with the robot kinematic model, the coordinate system transformation is performed to calculate the current attitude of the robot. The original 3D point cloud of the lidar is then pose-compensated and uniformly transformed to the world coordinate system. Based on the robot's current orientation, a local point cloud set in front of the robot is dynamically extracted from the global point cloud map scanned by LiDAR. The meaning of P here is inconsistent with the previous parameter; it should be changed to be consistent or a different parameter. , , The meaning of the parameters needs to be explained. Calculate the centroid of a local point cloud And construct the covariance matrix : ； right Perform singular value decomposition; the eigenvector corresponding to the smallest eigenvalue is the normal vector of the fitted plane. Then calculate the local slope. : ； Parameter meanings obstacle height This characterizes the geometric span of the plane within a local region. A local point cloud within a predetermined range in front of the robot is extracted, and the local support plane is fitted using the RANSAC algorithm as a reference. Then, the orthogonal distance from each sampling point in the point cloud set to the reference is calculated, and the maximum orthogonal distance within this region is recorded. and minimum value The difference is defined as the real-time obstacle height. : ； Roughness characterizes the degree of undulation in the ground. (Calculate the point cloud set.) The perpendicular distance of all points to the above-mentioned fitting plane Roughness is defined as the standard deviation of the distance. : ； The meaning of the parameters; Visual image data captured by the camera is input into the visual language model, which outputs terrain categories and probabilities. A hash mapping table of terrain semantics and physical parameters is pre-built in memory. The keys of the hash table are defined as the terrain semantic strings output by the VLM, including grassland, concrete, swamp, and land, and the values ​​are defined as values ​​containing the friction coefficient. After the visual language model outputs the terrain category, the friction coefficient corresponding to the terrain can be obtained from a table immediately. .

3. The method for smooth switching control of motion modes of a wheeled robot according to claim 2, characterized in that, The specific values ​​in the mapping table were obtained through offline experimental calibration, which involved dragging or slipping the wheeled robot on typical road surfaces, recording the corresponding maximum static friction and vertical vibration acceleration, and thus calculating the baseline values. Store in memory; After the multi-source sensors acquire data, the terrain feature vector can be obtained in real time based on the above calculation process.

4. The method for smooth switching control of motion modes of a wheeled robot according to claim 1, characterized in that, The method for mapping terrain feature vectors to joint weight matrices in step three includes: First, for each parameter in the terrain feature vector, a wheel-type attenuation / foot-type excitation function between 0 and 1 is designed. The evaluation functions for the four parameters of the feature vector are designed as follows: Use linear or smooth saturation mapping: ； The meanings of each parameter; please verify and supplement the meanings of other parameters. An increased gradient causes a sinusoidal increase in the driving torque required by the wheeled vehicle, which can easily lead to slippage or overturning. Therefore, a cosine attenuation model is used: ； Roughness causes continuous high-frequency vibrations, consuming wheel energy. An exponential decay model is adopted, where... The roughness sensitivity coefficient controls the degree of attenuation. ； When the height of the obstacle is close to the radius of the wheel At this point, the wheel absolutely cannot pass. A forced interception must be performed using a sigmoid function with soft boundaries. ； By integrating these four univariate evaluation factors, the overall confidence level or contribution of the current environment to wheeled sports can be calculated. : ； Calculate the overall confidence level or contribution of foot movements. for: ； Weighted null space control In the mathematical definition, the weight matrix Located in the denominator or inverse term of the quadratic optimization, meaning that the larger the weight value, the greater the penalty for the joint movement, and the more prone the joint is to locking. An improved sigmoid function is used to map the contribution to penalty weights in the joint augmented Jacobian matrix: The penalty weights for wheel-type and foot-type joints can be obtained as follows: ； in, The maximum penalty limit is set, with a value between 50 and 100. To switch sensitivity gain; Assuming a single leg has three joints (hip, knee, and ankle) and a wheel at its end, the joint weight matrix for that leg is constructed as follows: 。 5. The method for smooth switching control of motion modes of a wheeled robot according to claim 1, characterized in that, The specific process of constructing the control law based on the weighted pseudo-inverse in step four is as follows: Assume the robot has one leg with joints including hip lateral movement, hip pitch, and knee joint, and an end effector connected to a drive wheel. Define the leg joint variables as follows: Define the rolling joint variable of the wheel as Therefore, the complete generalized joint coordinates of this single-leg-wheel system are: ； Jacobian matrix Describes joint velocity With end effector linear velocity Linear mapping relationship between them: ； The velocity at the point of contact is the sum of the translation caused by the leg joint movement and the rolling translation caused by the wheel's rotation. Leg kinematic components. Considering only the linear velocity contribution of each leg joint motion to the wheel center, the wheel kinematic components are obtained by taking the partial derivative of the leg variables with respect to the robot's forward kinematics. Let the radius of the wheel be... The unit vector of the tangent direction in the plane containing the wheel is The contribution of wheel rolling to the linear velocity at the contact point is... The Jacobian column vector of the wheel is... By concatenating the two, we obtain the unified Jacobian matrix: ； Introducing the previously constructed diagonal weight matrix The formula for calculating the weighted pseudoinverse is: ； Based on the weighted pseudo-inverse matrix, an improved velocity-level null-space control law is constructed, and the calculation formula is as follows: ； in, For the desired joint velocity, For Jacobian matrices, For weighted pseudo-reverse, The expected speed of the main task, The weighted null space projection matrix ensures that the execution of subtasks is completely mapped within the null space of the main task, meaning that the movement of subtasks will not interfere with the main task. The expected joint velocity for the subtask. It is the identity matrix; To further eliminate sudden torque changes during mode switching and ensure a smooth response from the underlying motor controller, the aforementioned velocity-level control law is differentiated to construct an acceleration-level weighted zero-space control law: ； in, The main task expects acceleration. The time derivative of the Jacobian matrix. The desired acceleration for the subtask. The system can express the desired acceleration using the inverse dynamic equations. This is converted into the desired torque for each joint and sent to each joint actuator.

6. The method for smooth switching control of motion modes of a wheeled robot according to claim 1, characterized in that, The method for continuously differentiable processing of the control input in step five is as follows: To ensure the continuous differentiability of the control input (i.e., the continuity of the first derivative) and avoid abrupt acceleration changes, a cubic spline curve is used as the transition smoothing function. A normalized time variable is defined. ,in Constructing a smooth transition factor : ； The function satisfies And satisfy During the transition time window Internally, the continuous weight matrix actually distributed to the zero-space control law : ； When the time exceeds the transition window back, At this point, the system transitions to the new terrain completely and smoothly; The above time-domain filtered Substituting into the weighted null space control law in step four, the continuous and abruptly-free desired joint velocity is obtained. and desired joint acceleration .Will and Introducing the rigid body dynamics equations of a wheeled robot, we perform inverse dynamics feedforward calculations. The standard dynamics equations of the robot are expressed as: ； in The final desired motor drive torque vector is calculated as follows: Let be the symmetric positive definite inertia matrix of the robot. The matrix of Coriolis force and centrifugal force. Let the gravity compensation term vector be... This is a compensation term for joint friction. Due to the input At least have Continuity, calculated acceleration It is continuously distributed. Therefore, the final torque command derived through the mapping of the above dynamic equations... It is also continuous and smooth.

Images