A quadruped robot motion control method based on prior knowledge and its robot

Abstract

The present invention provides a method for controlling quadrupedal robot motion based on prior knowledge and the robot thereof, relating to the field of robot development technology, and comprising the following steps: S1: constructing the robot and a simulation environment; constructing a simulation environment for the quadrupedal robot, including a physical model of the quadrupedal robot, the physical properties of the surrounding environment, and sensors such as a depth camera, a lidar, and an IMU for the quadrupedal robot. This solution ultimately resolves the problems of weak robustness and stiff gait of quadrupedal robots in existing motion control methods, enabling the quadrupedal robot to achieve a more natural gait and improve its adaptability to rugged terrain. The results of this training method are less dependent on external environmental perception, and the quadrupedal robot can still make reasonable motion decisions even without terrain information. This method can significantly reduce training and debugging time, allowing the quadrupedal robot to be quickly deployed directly from a simulated environment to real-world applications.

Description

Technical Field

[0001] The present invention relates to the technical field of robot development, and in particular to a quadruped robot motion control method based on prior knowledge and the robot. Background Art

[0002] Improving the kinematic performance of quadruped robots plays a key role in their environmental adaptability, stability, and efficiency in complex terrains. This contributes to faster and more reliable mission execution and diverse applications, such as emergency rescue, logistics, and human-robot collaboration. Furthermore, optimizing kinematic performance promotes the development of bionics and artificial intelligence technologies, enhancing the autonomy and intelligence of quadruped robots and laying the foundation for the future advancement and widespread application of quadruped robot technology.

[0003] In existing methods, MPC has a trade-off between flexibility and safety margin during gait transitions, so the flexibility of quadruped robots will be limited, and this method is difficult to apply under the high-speed gait of quadruped robots; the high complexity of the WBC algorithm places high demands on hardware performance, which increases cost and complexity. In addition, the difficulties in actual operations such as determining task priorities, parameter adjustment and calibration also reduce the robustness and adaptability of WBC in certain scenarios; the disadvantage of the central pattern generator is the complexity of its parameter adjustment, which usually requires manual adjustment or genetic algorithm optimization, which makes the quadruped robot often only able to adapt to a single gait and lacks flexibility. Secondly, the central pattern generator relies on a coupling mechanism. Although it can generate stable gaits, these gaits are usually relatively rigid and unnatural.

[0004] Therefore, it is necessary to provide a quadruped robot motion control method and robot based on prior knowledge to solve the above technical problems. Summary of the Invention

[0005] The present invention provides a quadruped robot motion control method based on prior knowledge and the robot thereof, which solves the technical problems in the related art that the existing methods have weak robustness and generate stiff gaits.

[0006] To solve the above technical problems, the present invention provides a quadruped robot motion control method based on prior knowledge, which includes the following steps:

[0007] S1: Robot and simulation environment construction;

[0008] Build a quadruped robot simulation environment, including the quadruped robot's physical model, the physical properties of the surrounding environment, and the quadruped robot's depth camera, lidar, IMU and other sensors;

[0009] S2: skeleton recognition to obtain motion data;

[0010] The Microsoft Kinect depth camera captures depth images of the quadruped robot in various motion modes, including forward, backward, sideways movement, and steering, on flat terrain. A random decision forest classifier is used on the depth images to identify the various parts of the quadruped robot and classify each pixel. Based on the classification results, the mean shift algorithm is used to estimate the position of each joint of the quadruped robot. This algorithm determines the 3D position of the joints by finding the global center of mass at the location with the highest local density, satisfying the following requirements:

[0011]

[0012] Among them, x i is the coordinate of each point on the limbs and torso of the quadruped robot in the data set, x is the position of the current point, K is the Gaussian kernel function, which assigns a weight to each data point. The closer the point is, the higher the weight is. The formula is: σ is the bandwidth parameter, which determines the range of the kernel function. The algorithm moves the current point along the direction of the mean shift vector until it converges to the global center of mass on the current limb;

[0013] The skeleton model is optimized using Kalman filtering in each frame to obtain a smooth quadruped robot structure that satisfies:

[0014]

[0015] Where A is the state transfer matrix x k is the state vector of the joint In the vector, p0 is the initial position of the joint and v0 is the initial velocity;

[0016] The uncertainty of the predicted state using the covariance matrix P satisfies:

[0017] P k|k-1 =AP k-1 A T +Q

[0018] Where Q is the process noise covariance matrix, which is used to describe the randomness or uncertainty of the system model and initialize the covariance matrix P0;

[0019] Kalman gain K k Determines the balance between the predicted value and the observed value, and calculates the Kalman gain K k The formula is:

[0020] K k =P k|k-1 H T (HP k|k-1 H T +R) -1

[0021] Where H is the observation matrix, which is used to extract joint positions from the state vector, and R is the measurement noise covariance matrix, which is used to describe the reliability of the measurement;

[0022] According to the observed value z k Correct the predicted state to obtain the updated state vector x k satisfy:

[0023]

[0024] Among them, z k It is the measurement value at time step k, usually the observed position of the joint. The difference between the predicted value and the actual observed value is used to correct the predicted state, thereby reducing the position jitter caused by measurement noise and making the final estimated trajectory more stable and natural.

[0025] Through the above steps, a motion dataset is obtained, which contains the trajectories of various motion modes of the quadruped robot, such as forward, backward, sideways movement, and turning on flat terrain;

[0026] S3: Reinforcement learning framework design;

[0027] The robot's walking strategy is trained using reinforcement learning combined with an adversarial motion prior. The adversarial motion prior introduces a generative adversarial network to evaluate whether the robot's generated motion state is similar to the motion in a pre-defined reference motion dataset.

[0028] In a simulation environment, the Teacher-Student framework and the PPO algorithm are used to optimize the quadruped robot motion controller. A robust controller is learned to narrow the gap between the simulation model and the actual model, improve generalization ability, and generate an effective motion control strategy.

[0029] The teacher strategy has privileged information, which is usually only available in simulation and cannot be perceived in reality. During the training process, the teacher strategy can access rich state information, including proprioception information, privileged state and terrain information. The goal of the teacher strategy is to optimize the control strategy through reinforcement learning so that it can achieve stable walking on complex terrain. The teacher strategy not only relies on this rich state information, but also receives task rewards r task , style reward st and regularization reward r lt satisfy:

[0030] r t =r task +r st +r lt

[0031] Among them, rtask is the task reward, such as speed tracking or task completion, r st is a style reward based on adversarial motion priors that encourages the generation of natural gaits, r lt is a regularization reward used to constrain the smoothness and safety of the robot's motion;

[0032] The training goal of the generator is to make the robot obtain a higher style reward when performing actions, that is, to generate natural actions as close as possible to the data set. The style reward r st satisfy:

[0033] r st =max(0,1-0.25·(dscore t -1) 2 )

[0034] Among them, dscore t is the similarity score output by the discriminator. The higher the score, the more similar the current action of the quadruped robot is to the gait in the real data.

[0035] The optimization goal of the discriminator is to minimize the following loss function:

[0036]

[0037] Among them, (st,st+1)~D means that the samples come from the generated motion dataset D, (st,st+1)~A means that the samples come from the action sequence A generated by the generator. Is the score of the discriminator output, that is, the authenticity judgment of the state transition, output 1 indicates real movement, output -1 indicates it comes from the generator, The gradient penalty term constrains the gradient of the discriminator on the data samples to prevent the common mode collapse of the generative adversarial network during training.

[0038] Preferably, the student strategy is trained by imitating the behavior of the teacher strategy. Unlike the teacher strategy, the student strategy can only rely on the information obtained by the quadruped robot's body sensors and cannot access privileged state and terrain information. This means that the student strategy must be learned under a partially observable Markov decision process (POMDP). However, by learning the temporal correlation of historical data (such as changes in joint position and velocity over a period of time), the student strategy can infer the complexity of the current terrain or potential external interference. For this purpose, a long short-term memory network (LSTM) is used as the memory mechanism of the student strategy.

[0039] Preferably, the steps of training the student strategy include:

[0040] Action imitation: The student strategy learns the behavior of the teacher strategy in different states by imitating the action output of the teacher strategy;

[0041] Hidden state reconstruction: The student strategy attempts to reconstruct the teacher strategy’s hidden state representation using only ontological sensory information, which enables the student strategy to reproduce the teacher strategy’s behavior as accurately as possible without privileged information;

[0042] And in order to improve training efficiency, the low-level network of the student strategy will be initialized to the weights of the low-level network of the teacher strategy, thereby accelerating the training process.

[0043] A quadruped robot based on prior knowledge, comprising: a robot body, a switching and hiding mechanism, and a flipping mechanism;

[0044] The switching and hiding mechanism includes a mounting seat fixedly mounted on the top of the robot body, a mounting plate slidably connected to the inner side of the mounting seat, a bracket fixedly mounted on the right side of the mounting plate, a laser radar disposed on the top of the bracket, a gear plate fixedly mounted on the front side of the bottom of the mounting plate, a drive motor fixedly mounted on the top of the robot body, a gear fixedly mounted on the output shaft of the drive motor, and the gear meshing with the gear plate;

[0045] The flipping mechanism includes a rotating seat fixed to the top of the mounting plate, the top of the rotating seat is rotatably connected to a flip plate, the front and rear sides of the top of the mounting seat are fixed with guide plates, the two sides of the flip plate are slidably connected to the inner side of the guide plate, two arc plates are fixed on the right side of the flip plate, and a depth camera is fixed on the right side of the flip plate and on the side opposite to the two arc plates.

[0046] Preferably, the inner sides of the two guide plates are provided with guide grooves for use with the flap, and the rotation axis of the rotating seat is at a different position from the rotation axis of the flap. When the rotating seat drives the flap to move to the left, the bottom of the flap flips upward under the action of the guide groove.

[0047] Preferably, a horizontal plate is fixedly provided on the left side of the top of the two guide plates, and the inner side of the horizontal plate is threadedly connected to two threaded support rods.

[0048] Preferably, a support and protection mechanism is fixedly provided on the left side of the top of the mounting plate, and the support and protection mechanism includes two first rotating brackets fixedly provided on the left side of the top of the mounting plate, the inner sides of the two first rotating brackets are rotatably connected to connecting rods, the left ends of the two connecting rods are rotatably connected to second rotating brackets, the left sides of the two second rotating brackets are fixed with sliding brackets, the inner sides of the two sliding brackets are slidably connected to slide rails, the bottoms of the two slide rails are fixedly connected to the top of the mounting seat, the sides of the two sliding brackets separated from each other are fixed with protective brackets, the sides opposite to the two protective brackets are fixed with slides, and the two slides are respectively slidably connected to the inner sides of the two guide plates.

[0049] Preferably, a feedback mechanism is fixedly provided on the back side of the guide plate on the front side, and the feedback mechanism includes a sliding sleeve fixedly provided on the back side of the guide plate on the front side, a sliding rod is slidably connected to the inner side of the sliding sleeve, a spring is provided on the surface of the sliding rod, a contact plate is fixedly provided on the rear end of the sliding rod, a vertical plate is fixedly provided on the top of the slide plate on the front side, and a trigger pad is fixedly provided on the back side of the guide plate on the front side and on the inner side of the sliding sleeve.

[0050] Compared with related technologies, the quadruped robot motion control method based on prior knowledge and the robot provided by the present invention have the following beneficial effects:

[0051] This method can solve the problems of weak robustness and stiff gait of quadruped robots in existing motion control methods, enabling quadruped robots to obtain a more natural gait and improve their adaptability to rugged terrain; the results of this training method are less dependent on external environment perception. Even without terrain information, the quadruped robot can still make reasonable motion decisions. This method can significantly reduce training and debugging time, allowing quadruped robots to be quickly deployed directly from simulation environments to real-world applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.

[0053] Figure 1 A schematic diagram of the motion strategy generation method provided by the present invention;

[0054] Figure 2 A schematic diagram of the effect provided by the present invention;

[0055] Figure 3 The best structural diagram provided by the present invention;

[0056] Figure 4 for Figure 3 The structural diagram of the support and protection mechanism shown;

[0057] Figure 5 A schematic diagram of the structure of the switching and hiding mechanism provided by the present invention;

[0058] Figure 6 for Figure 5 A schematic structural diagram of a bottom view of the mounting base shown;

[0059] Figure 7 A schematic structural diagram of the flipping mechanism provided by the present invention;

[0060] Figure 8 for Figure 7 The structural diagram of the flap shown;

[0061] Figure 9 for Figure 7 The schematic diagram of the state in which the mounting plate drives the rotating seat to move to the left, causing the bottom of the flap to flip upward;

[0062] Figure 10 A schematic structural diagram of the support and protection mechanism provided by the present invention;

[0063] Figure 11 for Figure 10 The mounting plate shown moves to the left, and the two protective frames move away from each other under the action of the connecting rod;

[0064] Figure 12 This is a structural diagram of the feedback mechanism provided by the present invention.

[0065] Description of Figure Numbers:

[0066] 1. Robot body;

[0067] 2. Switching hidden mechanism; 21. Mounting seat; 22. Mounting plate; 23. Bracket; 24. LiDAR; 25. Gear plate; 26. Drive motor; 27. Gear;

[0068] 3. Flip mechanism; 31. Rotating seat; 32. Flip plate; 33. Guide plate; 34. Curved plate; 35. Depth camera;

[0069] 4. Horizontal plate; 5. Threaded support rod;

[0070] 6. Support and protection mechanism; 61. First rotating bracket; 62. Connecting rod; 63. Second rotating bracket; 64. Sliding bracket; 65. Slide rail; 66. Protection bracket; 67. Slide plate;

[0071] 7. Feedback mechanism; 71. Sliding sleeve; 72. Sliding rod; 73. Spring; 74. Contact plate; 75. Vertical plate; 76. Trigger pad.

[0072] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0073] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0074] The present invention provides a quadruped robot motion control method based on prior knowledge and the robot.

[0075] First embodiment:

[0076] See also Figure 1 and Figure 2 , a quadruped robot motion control method based on prior knowledge, comprising the following steps:

[0077] S1: Robot and simulation environment construction;

[0078] Build a quadruped robot simulation environment, including the quadruped robot's physical model, the physical properties of the surrounding environment, and the quadruped robot's depth camera, lidar, IMU and other sensors;

[0079] S2: skeleton recognition to obtain motion data;

[0080] The Microsoft Kinect depth camera captures depth images of the quadruped robot in various motion modes, including forward, backward, sideways movement, and steering, on flat terrain. A random decision forest classifier is used on the depth images to identify the various parts of the quadruped robot and classify each pixel. Based on the classification results, the mean shift algorithm is used to estimate the position of each joint of the quadruped robot. This algorithm determines the 3D position of the joints by finding the global center of mass at the location with the highest local density, satisfying the following requirements:

[0081]

[0082] Among them, x i is the coordinate of each point on the limbs and torso of the quadruped robot in the data set, x is the position of the current point, K is the Gaussian kernel function, which assigns a weight to each data point. The closer the point is, the higher the weight is. The formula is: σ is the bandwidth parameter, which determines the range of the kernel function. The algorithm moves the current point along the direction of the mean shift vector until it converges to the global center of mass on the current limb;

[0083] The skeleton model is optimized using Kalman filtering in each frame to obtain a smooth quadruped robot structure that satisfies:

[0084]

[0085] Where A is the state transfer matrix x k is the state vector of the joint In the vector, p0 is the initial position of the joint and v0 is the initial velocity;

[0086] The uncertainty of the predicted state using the covariance matrix P satisfies:

[0087] P k|k-1 =AP k-1 A T +Q

[0088] Where Q is the process noise covariance matrix, which is used to describe the randomness or uncertainty of the system model and initialize the covariance matrix P0;

[0089] Kalman gain K k Determines the balance between the predicted value and the observed value, and calculates the Kalman gain K k The formula is:

[0090] K k =P k|k-1 H T (HP k|k-1 H T +R) -1

[0091] Where H is the observation matrix, which is used to extract joint positions from the state vector, and R is the measurement noise covariance matrix, which is used to describe the reliability of the measurement;

[0092] According to the observed value z k Correct the predicted state to obtain the updated state vector x k satisfy:

[0093]

[0094] Among them, z k It is the measurement value at time step k, usually the observed position of the joint. The difference between the predicted value and the actual observed value is used to correct the predicted state, thereby reducing the position jitter caused by measurement noise and making the final estimated trajectory more stable and natural.

[0095] Through the above steps, a motion dataset is obtained, which contains the trajectories of various motion modes of the quadruped robot, such as forward, backward, sideways movement, and turning on flat terrain;

[0096] S3: Reinforcement learning framework design;

[0097] The robot's walking strategy is trained using reinforcement learning combined with an adversarial motion prior. The adversarial motion prior introduces a generative adversarial network to evaluate whether the robot's generated motion state is similar to the motion in a pre-defined reference motion dataset.

[0098] In a simulation environment, the Teacher-Student framework and the PPO algorithm are used to optimize the quadruped robot motion controller. A robust controller is learned to narrow the gap between the simulation model and the actual model, improve generalization ability, and generate an effective motion control strategy.

[0099] The teacher strategy has privileged information, which is usually only available in simulation and cannot be perceived in reality. During the training process, the teacher strategy can access rich state information, including proprioception information, privileged state and terrain information. The goal of the teacher strategy is to optimize the control strategy through reinforcement learning so that it can achieve stable walking on complex terrain. The teacher strategy not only relies on this rich state information, but also receives task rewards r task , style reward st and regularization reward r lt satisfy:

[0100] r t =r task +r st +r lt

[0101] Among them, r task is the task reward, such as speed tracking or task completion, r st is a style reward based on adversarial motion priors that encourages the generation of natural gaits, r lt is a regularization reward used to constrain the smoothness and safety of the robot's motion;

[0102] The training goal of the generator is to make the robot obtain a higher style reward when performing actions, that is, to generate natural actions as close as possible to the data set. The style reward r st satisfy:

[0103] r st =max(0,1-0.25·(dscore t -1) 2 )

[0104] Among them, dscore tis the similarity score output by the discriminator. The higher the score, the more similar the current action of the quadruped robot is to the gait in the real data.

[0105] The optimization goal of the discriminator is to minimize the following loss function:

[0106]

[0107] Among them, (st,st+1)~D means that the samples come from the generated motion dataset D, (st,st+1)~A means that the samples come from the action sequence A generated by the generator. Is the score of the discriminator output, that is, the authenticity judgment of the state transition, output 1 indicates real movement, output -1 indicates it comes from the generator, The gradient penalty term constrains the gradient of the discriminator on the data samples to prevent the common mode collapse of the generative adversarial network during training.

[0108] The student strategy is trained by imitating the behavior of the teacher strategy. Unlike the teacher strategy, the student strategy can only rely on information obtained from the quadruped robot's body sensors and cannot access privileged state and terrain information. This means that the student strategy must be learned under a partially observable Markov decision process (POMDP). However, by learning the temporal correlation of historical data (such as changes in joint position and velocity over a period of time), the student strategy can infer the complexity of the current terrain or potential external interference. To this end, a long short-term memory network (LSTM) is used as the memory mechanism of the student strategy.

[0109] The steps for training student strategies include:

[0110] Action imitation: The student strategy learns the behavior of the teacher strategy in different states by imitating the action output of the teacher strategy;

[0111] Hidden state reconstruction: The student strategy attempts to reconstruct the teacher strategy’s hidden state representation using only ontological sensory information, which enables the student strategy to reproduce the teacher strategy’s behavior as accurately as possible without privileged information;

[0112] And in order to improve training efficiency, the low-level network of the student strategy will be initialized to the weights of the low-level network of the teacher strategy, thereby accelerating the training process.

[0113] In this embodiment, compared with the existing method, this method uses a depth camera to capture the motion data of the quadruped robot, optimizes the trajectory through skeleton recognition technology and combined with Kalman filtering, and at the same time, uses a generative adversarial network (GAN) for adversarial training to generate a natural and smooth gait, thereby improving the motion performance of the quadruped robot; in the above process: on the one hand, first use skeleton recognition technology to extract motion data, and use Kalman filtering to optimize the trajectory, and then combine the generative adversarial network with reinforcement learning to generate a natural and smooth gait, and finally adopt a teacher-student strategy framework, the teacher strategy optimizes motion control, and the student strategy makes autonomous decisions in a partially observable environment, and ultimately robust motion control can be achieved. On the other hand, the results of this training method are less dependent on external environment perception. Even if the quadruped robot does not have terrain information, it can still make reasonable motion decisions, which can greatly reduce the training and debugging time, so that the quadruped robot can be quickly deployed directly from the simulation environment to real applications.

[0114] Second embodiment:

[0115] See also Figures 3 to 9 , a quadruped robot based on prior knowledge, comprising: a robot body 1, a switching and hiding mechanism 2 and a flipping mechanism 3;

[0116] The switching and hiding mechanism 2 includes a mounting seat 21 fixed to the top of the robot body 1, a mounting plate 22 is slidably connected to the inner side of the mounting seat 21, a bracket 23 is fixed to the right side of the mounting plate 22, a laser radar 24 is provided on the top of the bracket 23, a gear plate 25 is fixed to the front side of the bottom of the mounting plate 22, a drive motor 26 is fixed to the top of the robot body 1, and a gear 27 is fixed to the output shaft of the drive motor 26, and the gear 27 is meshed with the gear plate 25;

[0117] Please combine Figure 5 and Figure 6 : Start the drive motor 26, the drive motor 26 rotates to drive the gear 27 to rotate, the gear 27 then drives the gear plate 25 and the mounting plate 22 to move leftward, the mounting plate 22 moves leftward and drives the bracket 23 and the laser radar 24 to move leftward, thereby moving the laser radar 24 to the inner side of the mounting seat 21, thereby shielding the scanning range of the laser radar 24;

[0118] The flip mechanism 3 includes a rotating base 31 fixed to the top of the mounting plate 22, and a flap 32 is rotatably connected to the top of the rotating base 31. Guide plates 33 are fixed to the front and rear sides of the top of the mounting base 21. Both sides of the flap 32 are slidably connected to the inner side of the guide plate 33. Two arc-shaped plates 34 are fixed to the right side of the flap 32. A depth camera 35 is fixed to the right side of the flap 32 and on the side opposite to the two arc-shaped plates 34.

[0119] Please combine Figures 7 to 9 : When the mounting plate 22 moves to the left, it will also drive the rotating seat 31 to move to the left. During the movement of the rotating seat 31, it will drive the flap 32 to move to the left, so that both sides of the flap 32 slide to the left on the inner side of the guide plate 33. Under the action of the guide groove in the guide plate 33, when the mounting plate 22 moves to the leftmost position, the flap 32 flips 90 degrees counterclockwise, thereby adjusting the working angle of the depth camera 35 to the top.

[0120] The inner sides of the two guide plates 33 are provided with guide grooves for use with the flap 32. The rotation axis of the rotating seat 31 is at a different position from the rotation axis of the flap 32. When the rotating seat 31 drives the flap 32 to move to the left, the bottom of the flap 32 flips upward under the action of the guide groove.

[0121] A horizontal plate 4 is fixed on the left side of the top of the two guide plates 33, and two threaded support rods 5 are threadedly connected to the inner side of the horizontal plate 4;

[0122] Preferably, when the flap 32 rotates 90 degrees counterclockwise, the left side of the top of the flap 32 will contact the top of the threaded support rod 5. While the two threaded support rods 5 support the flap 32, excessive flipping of the flap 32 can be avoided.

[0123] In this embodiment, when the robot body 1 is trained using the quadruped robot motion control method, the drive motor 26 is started and the robot body 1 is adjusted to the training mode. The drive motor 26 rotates through the gear 27 to drive the gear plate 25 and the mounting plate 22 to move to the left, thereby moving the laser radar 24 to the inner side of the mounting seat 21, and the mounting seat 21 is used to shield the scanning range of the laser radar 24. During the movement, the mounting plate 22 will simultaneously drive the rotating seat 31 to move to the left. Under the action of the guide plate 33, the flap 32 drives the depth camera 35 to rotate 90 degrees counterclockwise, and the working angle of the depth camera 35 is adjusted to the top, thereby reducing the robot body 1's perception of the external environment, and observing the movement state of the robot body 1 when it has no terrain information, and whether the robot body 1 can make reasonable motion decisions.

[0124] Third embodiment:

[0125] See also Figure 4 、 Figures 10 to 12, a supporting and protective mechanism 6 is fixedly provided on the left side of the top of the mounting plate 22, and the supporting and protective mechanism 6 includes two first rotating brackets 61 fixedly provided on the left side of the top of the mounting plate 22, the inner sides of the two first rotating brackets 61 are rotatably connected to the connecting rods 62, the left ends of the two connecting rods 62 are rotatably connected to the second rotating brackets 63, and the left sides of the two second rotating brackets 63 are fixed with sliding brackets 64, the inner sides of the two sliding brackets 64 are slidably connected to the slide rails 65, and the bottoms of the two slide rails 65 are fixedly connected to the top of the mounting seat 21, and the separated sides of the two sliding brackets 64 are fixed with a protective bracket 66, and the opposite sides of the two protective brackets 66 are fixed with a slide plate 67, and the two slide plates 67 are respectively slidably connected to the inner sides of the two guide plates 33;

[0126] Please combine Figure 10 and Figure 11 : When the mounting plate 22 moves to the left and the robot body 1 is adjusted to the training mode, the movement of the mounting plate 22 will drive the two first rotating brackets 61 to move to the left. Under the action of the connecting rod 62, the two sliding frames 64 slide toward the side away from each other on the surface of the two slide rails 65. The movement of the two sliding frames 64 drives the two protective frames 66 to move toward the side away from each other. The movement of the two protective frames 66 causes the two slides 67 to slide toward the side away from each other on the inner side of the guide plate 33, thereby starting the protection state of the robot body 1.

[0127] A feedback mechanism 7 is fixedly provided on the back of the front guide plate 33. The feedback mechanism 7 includes a sliding sleeve 71 fixedly provided on the back of the front guide plate 33. A sliding rod 72 is slidably connected to the inner side of the sliding sleeve 71. A spring 73 is sleeved on the surface of the sliding rod 72. A contact plate 74 is fixedly provided on the rear end of the sliding rod 72. A vertical plate 75 is fixedly provided on the top of the front slide plate 67. A trigger pad 76 is fixedly provided on the back of the front guide plate 33 and inside the sliding sleeve 71.

[0128] Please combine Figure 4 and Figure 12 : When the two slides 67 move toward each other, the front slide 67 moves and drives the vertical plate 75 to move forward. The vertical plate 75 moves forward and drives the sliding rod 72 to slide forward inside the sliding sleeve 71, causing the spring 73 to contract. When the mounting plate 22 moves to the leftmost position, the front side of the sliding rod 72 will contact the trigger pad 76, thereby issuing a command information, and the staff can know that the working status of the robot body 1 has been switched.

[0129] In this embodiment, when the mounting plate 22 moves to the left, the mounting plate 22 causes the two sliding frames 64 to drive the two protective frames 66 to move to the side away from each other through the first rotating bracket 61, the connecting rod 62 and the second rotating bracket 63, thereby switching the robot body 1 to the training protection mode. During the switching process, the vertical plate 75 moves forward to drive the sliding rod 72 to slide forward inside the sliding sleeve 71. When the mounting plate 22 moves to the leftmost position, the front side of the sliding rod 72 will contact the trigger pad 76, and then issue a command information. The staff can know that the working status of the robot body 1 has been switched.

[0130] Please refer to the Figures 1 to 12 The quadruped robot motion control method based on prior knowledge and the working principle of the robot provided by the present invention are as follows:

[0131] Step S1, when the robot body 1 is trained using the quadruped robot motion control method, the drive motor 26 is first started. The drive motor 26 rotates to drive the gear 27 to rotate. The gear 27 drives the gear plate 25 and the mounting plate 22 to move leftward. The mounting plate 22 moves leftward, driving the bracket 23 and the laser radar 24 to move leftward, thereby moving the laser radar 24 to the inside of the mounting seat 21, shielding the scanning range of the laser radar 24;

[0132] In step S2, when the mounting plate 22 moves to the left, the rotating base 31 is also driven to move to the left. During the movement of the rotating base 31, the flap 32 is driven to the left, so that both sides of the flap 32 slide to the left on the inner side of the guide plate 33. Under the action of the guide groove in the guide plate 33, when the mounting plate 22 moves to the leftmost position, the flap 32 is flipped 90 degrees counterclockwise, adjusting the working angle of the depth camera 35 to the top.

[0133] In step S3, the mounting plate 22 moves to the left, which simultaneously drives the two first rotating brackets 61 to move to the left. Under the action of the connecting rod 62, the two sliding brackets 64 slide on the surfaces of the two slide rails 65 toward the separated side. The movement of the two sliding brackets 64 drives the two protective brackets 66 to move toward the separated side. The movement of the two protective brackets 66 further causes the two slide plates 67 to slide on the inner side of the guide plate 33 toward the separated side, starting the robot body 1 to the protective state. When the robot body 1 falls during training, the two protective brackets 66 are used to protect the robot body 1.

[0134] In step S4, when the two slides 67 move toward each other, the front slide 67 moves to drive the vertical plate 75 to move forward, and the vertical plate 75 moves forward to drive the sliding rod 72 to slide forward inside the sliding sleeve 71, causing the spring 73 to contract. When the mounting plate 22 moves to the leftmost position, the front side of the sliding rod 72 will contact the trigger pad 76, thereby issuing a command message. The staff can know that the working state of the robot body 1 has been switched. At this time, the left side of the rotating seat 31 will contact the right side of the two slides 67.

[0135] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. All equivalent structural transformations made by using the contents of the present invention description and drawings under the concept of the present invention, or direct / indirect application in other related technical fields are included in the patent protection scope of the present invention.

Claims

1. A quadruped robot motion control method based on prior knowledge, characterized in that: The following steps are involved: S1: Robot and simulation environment construction; Build a quadruped robot simulation environment, including the quadruped robot's physical model, the physical properties of the surrounding environment, and the quadruped robot's depth camera, lidar, and IMU sensor; S2: skeleton recognition to obtain motion data; The Microsoft Kinect depth camera captures depth images of the quadruped robot in four motion modes: forward, backward, sideways, and turning on flat terrain. A random decision forest classifier is used on the depth images to identify the various parts of the quadruped robot and classify each pixel. Based on the classification results, the mean shift algorithm is used to estimate the position of each joint of the quadruped robot. This algorithm determines the 3D position of the joints by finding the global center of mass at the location with the highest local density, satisfying the following requirements: ； in, are the coordinates of each point on the limbs and torso of the quadruped robot in the dataset, is the current point position, is a Gaussian kernel function, which assigns a weight to each data point. The closer the point is, the higher the weight is. Its formula is: , is the bandwidth parameter, which determines the range of the kernel function. The algorithm moves the current point along the direction of the mean shift vector until it converges to the global center of mass on the current limb; The skeleton model is optimized using Kalman filtering in each frame to obtain a smooth quadruped robot structure that satisfies: ； in, is the state transition matrix , is the state vector of the joint , in the vector is the initial position of the joint, is the initial velocity; Using the covariance matrix The uncertainty of the predicted state satisfies: ； in, is the process noise covariance matrix, which is used to describe the randomness or uncertainty of the system model. Initialize the covariance matrix ; Kalman gain Determines the balance between predicted and observed values ​​and calculates the Kalman gain The formula is: ； in, is the observation matrix, which is used to extract the joint positions from the state vector, is the measurement noise covariance matrix, which is used to describe the reliability of the measurement; According to the observed value Correct the predicted state to obtain the updated state vector satisfy: ； in, is in the time step The measured value is the observed position of the joint. The difference between the predicted value and the actual observed value is used to correct the predicted state, thereby reducing the position jitter caused by measurement noise and making the final estimated trajectory more stable and natural. Through the above steps, a motion dataset is obtained, which contains the trajectories of the quadruped robot moving forward, backward, sideways, and turning on flat terrain; S3: Reinforcement learning framework design; The robot's walking strategy is trained using reinforcement learning combined with an adversarial motion prior. The adversarial motion prior introduces a generative adversarial network to evaluate whether the robot's generated motion state is similar to the motion in a pre-defined reference motion dataset. In the simulation environment, the Framework and The algorithm optimizes the quadruped robot motion controller and learns a robust controller to narrow the gap between the simulation model and the actual model, improve generalization ability, and generate effective motion control strategies; During training, the teacher strategy is able to access rich state information, including proprioception information, privileged state, and terrain information. The goal of the teacher strategy is to optimize the control strategy through reinforcement learning so that the quadruped robot can achieve stable walking on complex terrain. The teacher strategy not only relies on this rich state information, but also receives task rewards. , Style Rewards and regularization reward satisfy: ； in, is a task reward, including speed tracking or task completion, It is a style reward based on adversarial motion priors that encourages the generation of natural gaits. is a regularization reward used to constrain the smoothness and safety of the robot's motion; The training goal of the generator is to make the robot obtain a higher style reward when performing actions, that is, to generate natural actions as close as possible to the data set. satisfy: ； Among them, dscore t is the similarity score output by the discriminator. The higher the score, the more similar the current action of the quadruped robot is to the gait in the real data. The optimization goal of the discriminator is to minimize the following loss function: ； in, Indicates that the sample comes from the generated motion dataset , Indicates that the sample comes from the action sequence generated by the generator , Is the score of the discriminator output, that is, the authenticity judgment of the state transition, output 1 indicates real movement, output -1 indicates it comes from the generator, It is a gradient penalty term that constrains the gradient of the discriminator on the data sample and prevents the common mode collapse of the generative adversarial network during training.

2. The quadruped robot motion control method based on prior knowledge according to claim 1 is characterized in that: A long short-term memory network (LSTM) is used as the memory mechanism of the student strategy. The training of the student strategy is achieved by imitating the behavior of the teacher strategy. Unlike the teacher strategy, the student strategy can only rely on the information obtained by the quadruped robot's body sensors and cannot access privileged state and terrain information.

3. The quadruped robot motion control method based on prior knowledge according to claim 2, characterized in that: The steps for training student strategies include: Action imitation: The student strategy learns the behavior of the teacher strategy in different states by imitating the action output of the teacher strategy; Hidden state reconstruction: The student strategy attempts to reconstruct the teacher strategy’s hidden state representation using only ontological sensory information, which enables the student strategy to reproduce the teacher strategy’s behavior as accurately as possible without privileged information; And in order to improve training efficiency, the low-level network of the student strategy will be initialized to the weights of the low-level network of the teacher strategy, thereby accelerating the training process.

4. A quadruped robot based on prior knowledge, characterized in that: The quadruped robot adopts the quadruped robot motion control method based on prior knowledge as described in any one of claims 1 to 3, comprising: a robot body, a switching and hiding mechanism, and a flipping mechanism; The switching and hiding mechanism includes a mounting seat fixedly mounted on the top of the robot body, a mounting plate slidably connected to the inner side of the mounting seat, a bracket fixedly mounted on the right side of the mounting plate, a laser radar disposed on the top of the bracket, a gear plate fixedly mounted on the front side of the bottom of the mounting plate, a drive motor fixedly mounted on the top of the robot body, a gear fixedly mounted on the output shaft of the drive motor, and the gear meshing with the gear plate; The flipping mechanism includes a rotating seat fixed to the top of the mounting plate, the top of the rotating seat is rotatably connected to a flip plate, the front and rear sides of the top of the mounting seat are fixed with guide plates, the two sides of the flip plate are slidably connected to the inner side of the guide plate, two arc plates are fixed on the right side of the flip plate, and a depth camera is fixed on the right side of the flip plate and on the side opposite to the two arc plates.

5. The quadruped robot based on prior knowledge according to claim 4, characterized in that: The inner sides of the two guide plates are provided with guide grooves for use with the flap. The rotation axis of the rotating seat is at a different position from the rotation axis of the flap. When the rotating seat drives the flap to move to the left, the bottom of the flap flips upward under the action of the guide groove.

6. The quadruped robot based on prior knowledge according to claim 4, characterized in that A horizontal plate is fixedly provided on the left side of the top of the two guide plates, and the inner side of the horizontal plate is threadedly connected to two threaded support rods.

7. The quadruped robot based on prior knowledge according to claim 4, characterized in that: A support and protection mechanism is fixedly provided on the left side of the top of the mounting plate, and the support and protection mechanism includes two first rotating brackets fixedly provided on the left side of the top of the mounting plate, the inner sides of the two first rotating brackets are rotatably connected to connecting rods, the left ends of the two connecting rods are rotatably connected to the second rotating brackets, and the left sides of the two second rotating brackets are fixed with sliding brackets, the inner sides of the two sliding brackets are slidably connected to slide rails, the bottoms of the two slide rails are fixedly connected to the top of the mounting seat, and the sides of the two sliding brackets separated from each other are fixed with protective brackets, and the sides opposite to the two protective brackets are fixed with slides, and the two slides are respectively slidably connected to the inner sides of the two guide plates.

8. The quadruped robot based on prior knowledge according to claim 7, characterized in that: A feedback mechanism is fixedly provided on the back side of the guide plate on the front side, and the feedback mechanism includes a sliding sleeve fixedly provided on the back side of the guide plate on the front side, a sliding rod is slidably connected to the inner side of the sliding sleeve, a spring is provided on the surface of the sliding rod, a contact plate is fixedly provided on the rear end of the sliding rod, a vertical plate is fixedly provided on the top of the slide plate on the front side, and a trigger pad is fixedly provided on the back side of the guide plate on the front side and on the inner side of the sliding sleeve.

Images