A quadruped robot positive obstacle climbing motion control method based on human motion capture reorientation

Abstract

The application discloses a quadruped robot positive obstacle climbing motion control method based on human motion capture redirection, which comprises the following steps: S1, collecting motion data of a human body in a four-limb supporting mode for climbing a positive obstacle; and S2, pre-processing the collected human motion data, and according to the human body contact state and the motion progress relative to the obstacle, etc. The motion data of the human body in the four-limb supporting mode for climbing the positive obstacle is collected, the rules of forelimb loading, trunk lifting, hindlimb stretching and gravity center shifting formed naturally by the human body are extracted, and the rules are taken as motion priori of the quadruped robot climbing, so that effective motion structure guidance is provided for a reinforcement learning strategy, invalid exploration behaviors can be effectively reduced, and training efficiency and stability are improved.

Description

Technical Field

[0001] This invention belongs to the fields of robot motion control, embodied intelligence and artificial intelligence, specifically referring to a method for controlling the motion of a quadruped robot climbing obstacles based on human motion capture redirection. Background Technology

[0002] Positive obstacle climbing is a crucial capability for quadruped robots in scenarios such as field inspection, emergency rescue, and navigating complex terrain. Positive obstacles refer to terrain structures that are significantly raised, protruding, or elevated relative to the robot's current standing plane or the surrounding ground. The robot needs to continuously perform actions such as approaching the obstacle, mounting with its front legs, raising its body, extending its hind legs, clearing the obstacle, and retracting its hind legs. This involves multiple contact state transitions during movement, and the movements of the body and four legs are coupled, making control significantly more difficult than ordinary walking.

[0003] Existing methods for positive obstacle climbing mainly employ manually designed actions, finite state machines, trajectory optimization, or reinforcement learning training. Manually designed actions and state machines rely heavily on empirical parameter tuning, making them difficult to adapt to robots of different sizes and obstacles. Trajectory optimization requires high model accuracy and computational power, and its adaptability is limited when model errors exist or the environment changes. Reinforcement learning, when training from scratch, lacks effective action priors, making the robot prone to ineffective explorations such as foreleg mounting failure, body collisions, unreasonable timing of hind leg push-offs, and instability and falls, resulting in long training cycles and unstable convergence.

[0004] When humans climb positive obstacles using all four limbs, they naturally exhibit movement patterns such as forelimb support, trunk lifting, hind limb extension, and weight transfer, providing an effective model for quadruped robots. However, significant differences exist between humans and quadruped robots in body proportions, joint topology, range of motion, and dynamic performance. Human motion capture data cannot be directly converted into robot joint commands; simple scaling may result in unreachable robot feet, joint over-limits, collisions, or instability. Therefore, a method for retargeting human motion capture and training robot motion strategies for positive obstacle climbing tasks is needed. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of current quadruped robots in imitating human climbing, such as unreachable feet, joint over-limits, collisions, or motion instability, and to provide a quadruped robot motion control method for positive obstacle climbing based on human motion capture retargeting.

[0006] To achieve the above objectives, the present invention employs the following technical solution: a method for controlling the forward obstacle climbing motion of a quadruped robot based on human motion capture retargeting, comprising the following steps: S1. Collect motion data of a human climbing a positive obstacle using a four-limb support method; S2. Preprocess the collected human motion data and divide the climbing process into multiple climbing stages based on the human contact state and the progress of movement relative to the obstacle. S3. Establish the mapping relationship between key parts of the human body and key parts of the quadruped robot; S4. Based on the human body size, quadruped robot size, robot joint range of motion, and obstacle size, redirect the human climbing motion into the quadruped robot's reference motion, and correct the motion feasibility of the redirection result. S5. Generate a reference trajectory for the quadruped robot to climb positive obstacles based on the redirection results; S6. Using the reference trajectory as a motion prior, construct a reinforcement learning motion strategy and train it to obtain a trained climbing control strategy. S7. Deploy the trained climbing control strategy to the quadruped robot. When the quadruped robot is actually running, based on the real-time acquired body state information and positive obstacle environment information, use the climbing control strategy to output control commands to control the quadruped robot to complete the positive obstacle climbing.

[0007] Furthermore, the climbing of the positive obstacle by the human body in step S1 using a four-limb support method refers to the human body using its hands and feet as support points to complete the climbing of the positive obstacle in a manner close to a four-legged movement pattern; the motion data includes the motion trajectory data of the human body trunk, limb joints, hands and feet, and contact state data between the hands and feet and the ground or obstacle; the positive obstacle includes at least one of steps, boxes, platforms or walls.

[0008] The preprocessing mentioned in step S2 refers to performing one or more of the following on the collected human motion data: coordinate unification, time synchronization, filtering and smoothing, and abnormal data processing; the climbing stage is at least divided into the approach and posture adjustment stage, the forelimb mounting stage, the trunk lifting and hindlimb extension stage, and the hindlimb retraction and posture stabilization stage.

[0009] The mapping relationship mentioned in step S3 includes: the human torso corresponds to the quadruped robot body, the human hands correspond to the quadruped robot's two forelegs, the human feet correspond to the quadruped robot's two hind legs, the human limbs' contact state corresponds to the quadruped robot's four-leg contact state, and the human climbing stage corresponds to the quadruped robot's climbing stage; the mapping refers to extracting and transferring one or more cross-morphological motion features from the torso movement, limb end movement, contact sequence, support relationship, and movement stage.

[0010] The relocation mentioned in step S4 refers to adjusting the size according to the proportional relationship between different body parts of the human body and the quadruped robot. This relocation includes: converting the human torso movement into the reference movement of the quadruped robot body, converting the human hands movement into the reference movement of the quadruped robot's two forelegs, and converting the human feet movement into the reference movement of the quadruped robot's two hind legs. The movement feasibility correction includes: limiting the range of motion of the quadruped robot's joints, adjusting the position of movements that are inaccessible to the feet, correcting collisions that may cause collisions with the body or legs, and adjusting the posture of movements that do not meet the support stability requirements.

[0011] The reference trajectory mentioned in step S5 includes one or more of the following information: the position, attitude, and velocity reference of the robot body; the motion trajectory of the four-legged end caps; the motion state of each joint; the contact state and contact sequence of the four legs; and climbing phase or motion phase information.

[0012] Step S6, which involves constructing and training a reinforcement learning motion strategy, includes the following steps: S61. Use the reference trajectory as a motion prior to guide the reinforcement learning strategy to learn the fuselage motion, foot motion, contact sequence and climbing phase; S62. During the training process, the reinforcement learning strategy dynamically adjusts the reference motion in the reference trajectory based on the real-time state, dynamic characteristics and obstacle conditions of the quadruped robot. S63. The reinforcement learning strategy directly outputs joint control commands, or outputs corrections relative to the reference action in the reference trajectory.

[0013] The training methods described in step S6 are: training using a unified motion strategy corresponding to the reference trajectory, training multiple motion sub-strategies trained separately according to the climbing stage, training in stages according to the climbing stage and then jointly optimizing, or training a unified conditional strategy with the climbing stage as input.

[0014] Furthermore, during the training process described in step S62, when the size of the obstacle changes, the reference trajectory changes accordingly with the size of the obstacle, and the reinforcement learning strategy learns the correction amount under the guidance of the changing reference trajectory, forming an adaptive climbing ability for obstacles of different sizes.

[0015] Meanwhile, during the training process described in step S62, one or more parameters among obstacle size, load, ground friction, sensor noise, and control delay also need to be randomized.

[0016] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention collects motion data of the human body climbing positive obstacles in a four-limb support manner, extracts the naturally formed laws of forelimb support, trunk lifting, hind limb extension and center of gravity transfer, and uses them as motion priors for quadruped robot climbing, providing effective action structure guidance for reinforcement learning strategies, which can effectively reduce ineffective exploration behavior and improve training efficiency and stability.

[0017] (2) This invention adopts cross-morphological semantic mapping, which does not rely on the one-to-one correspondence between human joints and robot joints, and makes size adaptation adjustments according to the proportional relationship between different body parts of the human body and the robot. At the same time, it makes feasibility corrections on the retargeting results in terms of joint range of motion, foot accessibility, collision prevention and support stability, thus avoiding the problem of motion transfer failure caused by the difference in morphology between the human body and the quadruped robot, and improving the applicability and safety of human demonstration to quadruped robot transfer.

[0018] (3) The present invention uses the reference trajectory generated by redirection as a motion prior to guide the training of reinforcement learning strategy. At the same time, it allows the strategy to dynamically adjust the reference action according to the robot's real-time state, dynamic characteristics and obstacle conditions, so that the robot can inherit the climbing rules in the human demonstration and form motion actions that conform to its own structure and dynamic characteristics.

[0019] (4) During the training process, the reference trajectory changes accordingly when the size of the obstacle changes. The strategy learns the correction amount under the guidance of the changing reference trajectory, forming an adaptive climbing ability for obstacles of different sizes. At the same time, the obstacle size, load, ground friction, sensor noise and control delay are randomized, which enhances the generalization ability of the strategy and the robustness of online operation. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0021] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0022] like Figure 1As shown in this embodiment, the quadruped robot obstacle climbing motion control method based on human motion capture redirection includes steps S1 to S7. This embodiment uses a quadruped robot with 12 active degrees of freedom (3 active joints per leg: hip lateral swing, hip forward swing, and knee joint) as an example for illustration. The quadruped robot has a body length of 0.6m, a standing height of 0.4m, a single leg length of 0.35m, and a total mass of approximately 25kg.

[0023] The specific implementation methods for each step are described in detail below.

[0024] S1. Collect motion data of the human body climbing a positive obstacle using a four-limb support method.

[0025] The step of climbing the obstacle using a four-limb support method refers to using an inertial motion capture device worn by the human body, with the hands and feet as support points, to complete the obstacle climb in a manner close to quadrupedal movement. The obstacle includes at least one of steps, boxes, platforms, or walls.

[0026] The motion data includes motion trajectory data of the human torso, limb joints, hands, and feet, as well as contact status data between the hands and feet and the ground or obstacles. Specifically, the motion data of the human torso includes position and posture data (6-dimensional, position x, y, z, and posture roll, pitch, and yaw angles); the motion data of the limb joints includes angle and angular velocity data of the four limb joints (shoulder, elbow, wrist, hip, knee, and ankle); the motion trajectory data of the hands and feet refers to the motion trajectory data of the hands and feet in three-dimensional space; the contact status data between the hands and feet and the ground or obstacles refers to data obtained through plantar pressure sensors or based on kinematic inference. A contact state is determined when the contact force exceeds a preset threshold of 5N; otherwise, it is determined to be an off-ground state.

[0027] Human climbing motion includes at least six consecutive phases: approaching the obstacle, forelimb mounting, trunk lifting, hind limb extension, body clearance, and hind limb retraction, covering the complete process from approaching the obstacle to complete clearance and regaining stability. Data from no fewer than 10 complete climbing cycles should be collected for each phase to ensure the sufficiency and representativeness of the motion data.

[0028] S2. Preprocess the collected human motion data and divide the climbing process into multiple climbing stages based on the human contact state and the progress of movement relative to the obstacle.

[0029] The preprocessing refers to performing one or more of the following on the collected human motion data: coordinate unification, time synchronization, filtering and smoothing, and abnormal data processing.

[0030] The aforementioned coordinate unification refers to unifying the data collected by each sensor into the same world coordinate system (with the center of the bottom edge of the obstacle as the origin, the x-axis pointing in the direction of robot movement, the y-axis pointing laterally, and the z-axis pointing vertically upward).

[0031] The time synchronization mentioned above refers to synchronizing the data from each sensor in time to ensure that the data from each channel are aligned in time (synchronization accuracy better than 1ms).

[0032] Filtering smoothing refers to using a fourth-order Butterworth low-pass filter with a cutoff frequency of 6Hz to smooth motion trajectory data and remove high-frequency noise.

[0033] Abnormal data processing refers to removing abnormal data points caused by sensor obstruction or communication interruption, and using cubic spline interpolation to complete missing data.

[0034] This embodiment divides the climbing process into four stages based on the contact status of the human body's hands and feet (whether the hands or feet are in contact with the ground or obstacle surface), the position of the torso (the height and horizontal position of the torso relative to the obstacle), and the progress of the human body's movement relative to the obstacle: (1) Approach and posture adjustment stage: The human body transitions from a standing posture to a four-limb support posture, with the hands and feet in contact with the ground, the torso leaning forward and beginning to move toward the obstacle.

[0035] (2) Forelimb mounting stage: The hands leave the ground and mount the upper surface of the obstacle. The torso begins to rise, while the feet remain in contact with the ground and provide support.

[0036] (3) Trunk lifting and hind limb extension stage: The trunk is further lifted, the legs extend to make the body's center of gravity pass over the edge of the obstacle, and the hands provide support on the upper surface of the obstacle.

[0037] (4) Hind limb retraction and posture stabilization stage: both feet leave the ground and are retracted to the upper surface of the obstacle, the body completely passes over the obstacle, and the limbs return to a stable supporting posture on the upper surface of the obstacle.

[0038] The stage information is appended to each frame of motion data in the form of a stage label sequence, which is used for subsequent reference trajectory generation, staged training, or motion strategy switching.

[0039] S3. Establish the mapping relationship between key parts of the human body and key parts of the quadruped robot.

[0040] In this step, the mapping relationship includes: the human torso corresponds to the quadruped robot body; the human hands correspond to the quadruped robot's two forelegs; the human feet correspond to the quadruped robot's two hind legs; the human limbs' contact state corresponds to the quadruped robot's four-leg contact state; and the human climbing stage corresponds to the quadruped robot's climbing stage.

[0041] The mapping refers to extracting and transferring one or more cross-morphological motion features from trunk movements, limb extremity movements, contact sequences, support relationships, and movement phases. This embodiment completely abandons the traditional inertial thinking that requires a one-to-one correspondence between human joints and robot joints. Instead, it extracts and transfers cross-morphological motion features such as trunk movements, limb extremity movements, contact sequences, support relationships, and movement phases. That is, this embodiment transfers high-level motion patterns such as "how the trunk is raised," "when and where the forelimbs (hands / forefeet) mount," "when and where the hind limbs (feet / hind feet) push off," and "how the center of gravity shifts," rather than specific joint angle sequences. This semantic-level mapping method can effectively avoid the problem of directly copying joint movements caused by the difference in joint topology between humans and robots.

[0042] S4. Based on the human body size, quadruped robot size, robot joint range of motion, and obstacle size, redirect the human climbing motion into the quadruped robot's reference motion, and correct the motion feasibility of the redirection result.

[0043] The retargeting described herein is not a simple proportional scaling of human body movements, but rather a dimensional adaptation adjustment based on the proportional relationship between different body parts of the human body and the quadruped robot. This retargeting includes: converting human torso movements into reference movements for the quadruped robot's body, converting human hand movements into reference movements for the quadruped robot's forelegs, and converting human foot movements into reference movements for the quadruped robot's hind legs.

[0044] Converting human torso motion into reference motion for the quadruped robot body, known as torso motion redirection, specifically refers to transforming the position and orientation trajectory of the human torso in the world coordinate system into the desired position and orientation trajectory of the quadruped robot body. During the conversion, the robot body dimensions (0.6m long, 0.3m wide) and center of mass height are considered. By calculating the displacement and rotation angles of the human torso during climbing, these are mapped to the robot body motion proportionally to the robot's dimensions.

[0045] Converting human hand movements into reference movements for the quadruped robot's forelegs, known as foreleg motion redirection, specifically involves transforming the motion trajectories of the human hand ends into the desired motion trajectories of the robot's forelegs. The contact objects (ground or obstacle surface) and contact timing of the hands are preserved and mapped to the contact objects and contact timing of the forelegs. The mapping formula from hand position to foreleg position is: Foreleg position = Human hand position × (Robot foreleg length / Human upper limb length) × Direction adjustment coefficient.

[0046] The process of converting human bipedal motion into reference motion for the hind legs of a quadruped robot, known as hind leg motion redirection, specifically involves transforming the motion trajectories of the human bipedal ends into the desired motion trajectories of the robot's hind legs. The contact objects and contact timing of the bipedal ends are preserved and mapped to the contact objects and contact timing of the hind legs.

[0047] The aforementioned motion feasibility correction includes: limiting the range of motion of the quadruped robot joints, adjusting the position of movements that are inaccessible to the feet, correcting collisions that may cause collisions with the robot body or legs, and adjusting the posture of movements that do not meet the support stability requirements.

[0048] Among them, the range of motion correction of joints refers to checking whether the angles of each joint exceed the hardware limits (hip joint lateral swing: ±30°, hip joint forward swing: -45° to +90°, knee joint: -150° to -30°), and correcting the excessive movements to ensure that the joint angles are always within the allowable range.

[0049] Adjusting the position of movements that are inaccessible to the feet means checking whether each foot position is within the kinematically accessible space of the leg (a spherical shell area with a radius of 0.35m and the accessible range of the feet as the origin), and adjusting the position of inaccessible foot positions to make them fall into the accessible space.

[0050] Collision avoidance correction for actions that may result in fuselage or leg collisions refers to: checking whether there is any interference collision between the fuselage and leg components and the obstacle surface (simplifying the fuselage and legs as an enveloping sphere with a radius of 0.05m, and the obstacle surface as a plane), adjusting the collision avoidance for actions that may result in collisions, and ensuring that the minimum distance between the fuselage and legs and the obstacle surface is greater than the safety threshold of 0.02m.

[0051] Adjusting the posture of actions that do not meet the support stability requirements means checking whether the robot's center of gravity projection falls inside the support polygon, adjusting postures that do not meet the stability conditions, and ensuring support stability (stability margin greater than 0.02m).

[0052] S5. Generate a reference trajectory for the quadruped robot to climb obstacles based on the redirection results.

[0053] The reference trajectory mentioned in this step includes one or more of the following information: the position, attitude, and velocity reference of the robot body; the motion trajectory of the four legs; the motion state of each joint; the contact state and contact sequence of the four legs; and climbing phase or motion phase information.

[0054] The robot's position, attitude, and speed references specifically include the expected three-dimensional position (accuracy 0.01m), expected attitude angles (roll angle, pitch angle, yaw angle, accuracy 0.5°), and expected speed (accuracy 0.01m / s) of the robot at each moment, with a time resolution of 20ms (consistent with the robot control cycle).

[0055] The trajectory of the four feet refers to the expected three-dimensional position of the four feet at each moment (accuracy 0.005m).

[0056] The motion state of each joint refers to the expected angle (accuracy 0.1°) and angular velocity (accuracy 0.5° / s) of 12 joints calculated by inverse kinematics from the foot position and fuselage attitude.

[0057] The quadrupedal contact state and contact sequence refer to the expected contact state (contact / off-ground, binary quantity) of each foot at each moment, and the order in which the feet establish contact with the ground or obstacles.

[0058] Climbing phase or motion phase information refers to the time boundaries and transition conditions of each climbing phase, as well as the climbing phase identifier of each moment (0 to 3 correspond to four phases respectively).

[0059] After generating the reference trajectory for the quadruped robot to climb obstacles, the trajectory needs to be checked for motion feasibility again to ensure that it meets the requirements of the quadruped robot's joint range of motion, foot accessibility, collision avoidance between the body and legs, contact stability, and continuous motion.

[0060] S6. Using the reference trajectory as a motion prior, construct a reinforcement learning motion strategy and train it to obtain the trained climbing control strategy.

[0061] The construction of the reinforcement learning motion policy mentioned above refers to constructing a reinforcement learning policy network based on the Actor-Critic architecture. The observation space of this policy network contains the following three parts of information: The body state information includes fuselage attitude angles (roll angle, pitch angle, yaw angle, 3D), fuselage angular velocity (3D), fuselage linear velocity (3D), 12 joint angles (12D), 12 joint angular velocities (12D), 4 foot positions (12D), and 4 foot contact states (4D), for a total of 49 dimensions.

[0062] The obstacle environment information includes obstacle height (1D), obstacle width (1D), obstacle edge position (2D), relative distance between the robot and the obstacle (1D), and relative azimuth angle (1D), for a total of 6 dimensions.

[0063] The reference trajectory information includes the current position of each foot target (12 dimensions), the current attitude of the fuselage target (3 dimensions), the current contact state of the target (4 dimensions), and the sequence of foot target positions for the next 5 moments (60 dimensions), for a total of 79 dimensions.

[0064] The total number of dimensions is 134. The data for each dimension is input into the policy network after being normalized.

[0065] In this embodiment, the policy network adopts a multilayer perceptron (MLP) structure, which sequentially includes an input layer (134-dimensional), three hidden layers (with 256, 256, and 128 neurons respectively, and ReLU activation function), and an output layer (12-dimensional, tanh activation function). The policy network outputs a correction amount relative to the reference action in the reference trajectory. The actual control command is generated by superimposing the reference action and the correction amount, that is: joint angle command = reference joint angle + correction amount output by the policy network.

[0066] The policy network is trained using the Proximal Policy Optimization (PPO) algorithm, with the following hyperparameter settings: learning rate 3 × 10⁻⁶. -4 The discount factor γ=0.99, the GAE coefficient λ=0.95, the clipping range ε=0.2, and 2048 time steps of interaction data are collected in each round. Ten rounds of mini-batch updates are performed, with a batch size of 256.

[0067] After constructing the reinforcement learning motion strategy, the strategy needs to be trained. In this embodiment, the following methods are used for training: training a unified motion strategy corresponding to the reference trajectory, training multiple motion sub-strategies trained separately according to the climbing stage, training in stages according to the climbing stage and then jointly optimizing, or training a unified conditional strategy with the climbing stage as input.

[0068] The training includes the following steps: S61. Using the reference trajectory as a motion prior, guide the reinforcement learning strategy to learn the body motion, foot motion, contact sequence, and climbing phase.

[0069] S62. During the training process, the reinforcement learning strategy dynamically adjusts the reference motion in the reference trajectory based on the real-time state, dynamic characteristics and obstacle conditions of the quadruped robot.

[0070] S63. The reinforcement learning strategy directly outputs joint control commands, or outputs corrections relative to the reference action in the reference trajectory.

[0071] It should be noted that during the training process described in step S62, when the obstacle size changes, the reference trajectory changes accordingly, and the reinforcement learning strategy learns correction values ​​under the guidance of the changing reference trajectory, forming an adaptive climbing ability for obstacles of different sizes. Simultaneously, one or more parameters among obstacle size, load, ground friction, sensor noise, and control delay are randomized.

[0072] Randomizing obstacle dimensions means that the obstacle height varies randomly within ±30% of the nominal value, and the obstacle width varies randomly within ±20% of the nominal value.

[0073] Load randomization refers to randomly applying an additional load of 0 to 2 kg to the torso of a quadruped robot.

[0074] Randomization of ground friction refers to uniformly and randomly sampling the ground friction coefficient within the range of 0.5 to 1.2.

[0075] Sensor noise randomization refers to applying Gaussian noise with a mean of zero and a standard deviation equal to the nominal noise level of the corresponding sensor to the observation vector (standard deviation of joint angle noise is 0.01 rad, and standard deviation of attitude angle noise is 0.005 rad).

[0076] Control delay randomization refers to randomly applying a delay of 0 to 2 control cycles (0 to 40 ms).

[0077] The aforementioned randomization process not only enhances the generalization ability and robustness of the strategy during online operation, but more importantly, it transforms static human demonstration data into dynamic priors covering various obstacle scenarios through a combination of redirection and randomized training. In this embodiment, the reinforcement learning strategy converges its network parameters and achieves a stable climbing success rate of over 90% after approximately 5 million training steps. At this point, the network parameters are exported and solidified, yielding the trained climbing control strategy.

[0078] S7. Deploy the trained climbing control strategy to the quadruped robot. When the quadruped robot is actually running, based on the real-time acquired body state information and positive obstacle environment information, use the climbing control strategy to output control commands to control the quadruped robot to complete the positive obstacle climbing.

[0079] This step involves deploying the trained climbing control strategy onto the onboard computer of the real quadruped robot.

[0080] The deployment method in this embodiment is as follows: the trained policy network parameters (weights and biases of the Actor network) are loaded into the inference framework of the quadruped robot's onboard computer (using ONNX Runtime as the inference engine). The inference frequency of the policy network is set to 50Hz (i.e., forward inference is performed once every 20ms), which is consistent with the control frequency of the underlying servo controller.

[0081] Its online operation process is as follows: (1) Real-time acquisition of body status information (body posture, joint angle, foot contact status, etc.) and positive obstacle environment information (obstacle height, width, edge position, relative distance, etc.) through airborne sensing devices (front depth camera, IMU, joint encoder, foot force sensor, etc.); (2) Based on the acquired obstacle information, select a matching reference trajectory from the pre-generated reference trajectory library (the reference trajectory library contains pre-generated reference trajectories for different obstacle sizes (height 0.15m to 0.45m, width 0.3m to 0.8m discretized with a step size of 0.05m), or generate a reference trajectory adapted to the actual obstacle size online by linear interpolation. (3) The robot’s current state (49-dimensional body state), obstacle information (6-dimensional environment information) and reference trajectory information (79-dimensional) are concatenated into a 134-dimensional multidimensional observation vector according to the format defined in S61 and then normalized. (4) Input the normalized observation vector into the deployed policy network and perform forward inference calculation to obtain the 12-dimensional action correction amount; (5) The correction amount is superimposed on the reference motion in the reference trajectory to generate the final control command: joint angle command = reference joint angle + correction amount; (6) Send control commands to the underlying servo controller (using a PD controller, with proportional gain Kp=50 and differential gain Kd=2) to drive the actuators of each joint of the robot to complete the corresponding actions.

[0082] During execution, the system adjusts the execution speed, movement phase, or strategy switching timing of the reference trajectory online based on the actual foot contact state and movement progress. Specifically: when there is a deviation between the actual foot placement position and the reference trajectory (deviation exceeding 0.02m), the strategy network outputs a corresponding correction for compensation; when the foot contact state does not match expectations (e.g., expected contact but actual non-contact), the system advances or delays the movement phase switching based on the actual contact state; when the quadruped robot's movement progress lags behind the reference trajectory (e.g., the actual body position lags behind the reference trajectory by more than 0.05m), the system appropriately slows down the execution speed of the reference trajectory, waiting for the robot to catch up before continuing.

[0083] When the quadruped robot's body crosses the upper surface of the obstacle, at least three feet are stably in contact with the ground (contact force greater than 5N), and the body posture angle is within the preset safety threshold range (roll angle |φ|<15°, pitch angle |θ|<15°), the obstacle climbing task is considered complete.

[0084] As described above, the present invention can be well implemented.

Claims

1. A method for controlling the forward obstacle climbing motion of a quadruped robot based on human motion capture retargeting, characterized in that, Includes the following steps: S1. Collect motion data of a human climbing a positive obstacle using a four-limb support method; S2. Preprocess the collected human motion data and divide the climbing process into multiple climbing stages based on the human contact state and the progress of movement relative to the obstacle. S3. Establish the mapping relationship between key parts of the human body and key parts of the quadruped robot; S4. Based on the human body size, quadruped robot size, robot joint range of motion, and obstacle size, redirect the human climbing motion into the quadruped robot's reference motion, and correct the motion feasibility of the redirection result. S5. Generate a reference trajectory for the quadruped robot to climb positive obstacles based on the redirection results; S6. Using the reference trajectory as a motion prior, construct a reinforcement learning motion strategy and train it to obtain a trained climbing control strategy. S7. Deploy the trained climbing control strategy to the quadruped robot. When the quadruped robot is actually running, based on the real-time acquired body state information and positive obstacle environment information, use the climbing control strategy to output control commands to control the quadruped robot to complete the positive obstacle climbing.

2. The method for controlling the forward obstacle climbing motion of a quadruped robot based on human motion capture redirection according to claim 1, characterized in that, The climbing of a positive obstacle by the human body in step S1 using a four-limb support method refers to the human body using its hands and feet as support points to complete the climbing of the positive obstacle in a manner close to a four-legged movement pattern. The motion data includes motion trajectory data of the human torso, limb joints, hands and feet, as well as contact status data between the hands and feet and the ground or obstacles; The positive obstacle includes at least one of steps, boxes, platforms, or wall sills.

3. The method for controlling the forward obstacle climbing motion of a quadruped robot based on human motion capture redirection according to claim 1, characterized in that, The preprocessing mentioned in step S2 refers to performing one or more of the following on the collected human motion data: coordinate unification, time synchronization, filtering and smoothing, and abnormal data processing. The climbing phase is divided into at least the approach and posture adjustment phase, the forelimb mounting phase, the trunk lifting and hindlimb extension phase, and the hindlimb retraction and posture stabilization phase.

4. The method for controlling the forward obstacle climbing motion of a quadruped robot based on human motion capture redirection according to claim 1, characterized in that, The mapping relationship mentioned in step S3 includes: the human torso corresponds to the quadruped robot body, the human hands correspond to the quadruped robot's two forelegs, the human feet correspond to the quadruped robot's two hind legs, the human limbs' contact state corresponds to the quadruped robot's four-leg contact state, and the human climbing stage corresponds to the quadruped robot's climbing stage; the mapping refers to extracting and transferring one or more cross-morphological motion features from the torso movement, limb end movement, contact sequence, support relationship, and movement stage.

5. The method for controlling the forward obstacle climbing motion of a quadruped robot based on human motion capture redirection according to claim 1, characterized in that, The redirection mentioned in step S4 refers to adjusting the size according to the proportional relationship between different body parts of the human body and the quadruped robot. The redirection includes: converting the human torso movement into the reference movement of the quadruped robot body, converting the human hands movement into the reference movement of the quadruped robot's two forelegs, and converting the human feet movement into the reference movement of the quadruped robot's two hind legs. The motion feasibility correction includes: limiting the range of motion of the quadruped robot joints, adjusting the position of movements that are inaccessible to the feet, correcting collisions that may cause collisions with the robot body or legs, and adjusting the posture of movements that do not meet the support stability requirements.

6. The method for controlling the forward obstacle climbing motion of a quadruped robot based on human motion capture redirection according to claim 1, characterized in that, The reference trajectory mentioned in step S5 includes one or more of the following information: the position, attitude, and velocity reference of the robot body; the motion trajectory of the four-legged end caps; the motion state of each joint; the contact state and contact sequence of the four legs; and climbing phase or motion phase information.

7. The method for controlling the forward obstacle climbing motion of a quadruped robot based on human motion capture redirection according to claim 1, characterized in that, Step S6, which involves constructing and training a reinforcement learning motion strategy, includes the following steps: S61. Using the reference trajectory as a motion prior, guide the reinforcement learning strategy to learn the fuselage motion, foot motion, contact sequence and climbing phase; S62. During the training process, the reinforcement learning strategy dynamically adjusts the reference motion in the reference trajectory based on the real-time state, dynamic characteristics and obstacle conditions of the quadruped robot. S63. Reinforcement learning strategies directly output joint control commands, or output corrections relative to reference actions in a reference trajectory.

8. The method for controlling the forward obstacle climbing motion of a quadruped robot based on human motion capture redirection according to claim 7, characterized in that, The training methods described in step S6 are: training using a unified motion strategy corresponding to the reference trajectory, training multiple motion sub-strategies trained separately according to the climbing stage, training in stages according to the climbing stage and then jointly optimizing, or training a unified conditional strategy with the climbing stage as input.

9. The method for controlling the forward obstacle climbing motion of a quadruped robot based on human motion capture redirection according to claim 8, characterized in that, During the training process described in step S62, when the size of the obstacle changes, the reference trajectory changes accordingly with the size of the obstacle, and the reinforcement learning strategy learns the correction amount under the guidance of the changing reference trajectory, forming an adaptive climbing ability for obstacles of different sizes.

10. The method for controlling the forward obstacle climbing motion of a quadruped robot based on human motion capture redirection according to claim 9, characterized in that, During the training process described in step S62, one or more parameters among obstacle size, load, ground friction, sensor noise, and control delay are also randomized.

Images