A tactile-based robot foot ground compliance control method and control system

Abstract

The present invention belongs to the technical field of footed robots, and specifically relates to a tactile-based robot foot ground compliance control method. The present invention includes the following steps: S1. Building a foot motion simulation platform; S2. Setting a mechanism for simulating the movement of the foot robot's legs and feet and a plantar pressure sensing system on the simulation platform; S3. Collecting and analyzing tactile information during the actual plantar compliance process of the plantar robot; S4. Constructing a control algorithm system based on deep reinforcement learning; S5. Establishing a mapping relationship between the virtual experiment and the real environment. By setting a number of simulated small balls to simulate the sensor points on the plantar of the actual plantar robot, the dynamic changes of tactile information during the movement of the robot arm are characterized. Then, the simulated control data is obtained through the control algorithm system based on deep reinforcement learning. The actual plantar robot is dynamically adjusted to its compliance posture through the mapping relationship, thereby achieving the completion of the plantar compliance task between the plantar robot and the ground, and autonomously adapting to the actual environment based on tactile information.

Description

Technical Field

[0001] The present invention belongs to the technical field of footed robots, and in particular relates to a tactile-based robot foot ground compliance control method and control system. Background Art

[0002] Legged robots have attracted much attention due to their application potential in emergency rescue, military operations, daily life assistance for disabled elderly people, as well as services and industrial production. Existing legged robots can use passive ankle joints to achieve compliance with the ground, but the passive joints have no rigidity and cannot achieve active control. Although they can passively comply with the ground, when there is a risk of overall instability, the balance of the robot as a whole cannot be controlled through ankle joint operation, which is an obvious shortcoming.

[0003] For example, to improve the leg control accuracy of a legged robot walking on soft surfaces, Zhou et al. proposed a method that combines force control with ground learning. This method first identifies the ground type and properties, then adjusts the leg movements through force control to ensure that the robot's foot force follows a preset reference force, enabling the robot to walk smoothly on soft surfaces. However, to address the problem of ground compliance of the robot's feet, this method overly relies on the robot's own terrain detection results and lacks information such as contact pressure distribution, contact point location, and ground hardness. This results in the robot's lack of adaptability and stability on some more challenging terrains. Summary of the Invention

[0004] The purpose of the present invention is to overcome the defects of insufficient foot adaptability and stability in the prior art, and to provide a tactile-based robot foot ground compliance control method and control system that can perceive the ground according to tactile information of the sole of the foot and autonomously adapt to the ground.

[0005] The technical solution adopted by the present invention to solve its technical problem is:

[0006] As a first aspect, a tactile-based robot foot ground compliance control method includes the following steps:

[0007] S1. Build a foot motion simulation platform;

[0008] S2. A mechanism for simulating the movement of the robot leg foot and a plantar pressure sensing system is provided on the simulation platform;

[0009] S3. Collect and analyze tactile information during the actual foot robot plantar compliance process;

[0010] S4. Build a control algorithm system based on deep reinforcement learning to calculate and obtain simulated control data based on the collected and analyzed tactile information;

[0011] S5. Establish a mapping relationship between the virtual experiment and the real environment, that is, map the simulated control data to the actual foot robot and adjust the actual foot robot to conform to the posture;

[0012] The simulation platform includes a simulated ground, a simulated foot robot and several simulated balls. The simulated balls are arranged and distributed on the soles of the simulated robots, corresponding to the sensor points in the sole pressure sensing system. The dynamic changes of tactile information in the actual movement of the robotic arm are represented according to the number of detected collision sensor points.

[0013] Specifically, the diameter of the simulated ball is 3-7 mm; the plantar pressure sensing system includes a tactile acquisition module for collecting tactile information of the actual plantar of the foot robot and an analysis and calculation module corresponding to the simulated ball, and the tactile acquisition module includes 128 sensor points arrayed on the plantar of the actual foot robot.

[0014] Specifically, the simulated ground is set as a double-layer simulation, including an upper ground layer, a soft material layer and a lower reference layer constructed sequentially from top to bottom;

[0015] When the sensing point of the robot foot touches the space between the upper ground layer and the lower reference layer, it is recognized as a collision and the compression of the soft material layer is simulated.

[0016] Specifically, the control algorithm system adopts a deep reinforcement learning algorithm based on the combination of value function and policy gradient;

[0017] The deep reinforcement learning algorithm includes the following steps:

[0018] S401. Construct a reward and punishment model;

[0019] S402. Establish a state space; that is, the number of collisions between the sensor point and the upper and lower planes, the multiple joint angles of the manipulator, and the vertical coordinate Z of the center of the manipulator end are incorporated into the state space. The state space formula is:

[0020] S=[θ1,θ2,θ3,…θn,up_collision,bottom_collison,Z];

[0021] S=[θ1,θ2,θ3,θ4,θ5,θ6,up_collision,bottom_collison,z];

[0022] Among them, up_collision represents the number of collisions between the sensor point and the upper plane, bottom_collision represents the number of collisions between the sensor point and the lower plane, and θ1-θn represents multiple joint angles of the robot arm;

[0023] S403. Construct an action space; that is, define a multidimensional space that only includes the changes in multiple joint angles of the robotic arm. The corresponding formula of the action space A is:

[0024] A=[Δθ1, Δθ2, Δθ3,…, Δθn];

[0025] A=[Δθ1,Δθ2,Δθ3,Δθ4,Δθ5,Δθ6];

[0026] Among them, Δθ1, Δθ2, Δθ3, ..., Δθn are the angle changes of each joint of the robotic arm corresponding to each dimension in the next action cycle.

[0027] Specifically, the reward and punishment model includes:

[0028] Base reward: When a sensor point is in contact only with the upper ground layer, a positive reward is assigned to the sensor point:

[0029] baseReward = up_collision;

[0030] Where baseReward is the base reward, and up_collision is the number of sensor points in contact with the upper ground.

[0031] Penalty mechanism: If a sensor point contacts the underlying reference surface, a negative reward is assigned to the sensor point:

[0032] negativeReward=-bottom_collison;

[0033] Where negativeReward is the penalty part, bottom_collision is the number of sensor points in contact with the reference plane;

[0034] Sparse rewards:

[0035] downReward=2.5×(pre_z-z);

[0036] Where z represents the vertical coordinate of the end of the manipulator, pre_z represents the vertical coordinate of the end of the manipulator at the beginning of each training cycle, and the unit is mm. If the foot moves in the negative direction of the z axis, a positive reward is given; on the contrary, if the foot moves in the positive direction of the z axis, it is considered a non-target behavior and a negative reward is given.

[0037] SuccessReward reward mechanism: When the number of sensor points where the robot's foot detects a collision reaches 100% of the total number of sensor points, that is, the robot's foot is in full contact with the ground, the task is considered successfully completed and the SuccessReward value is set to +10. If this standard is not met, the reward value remains 0.

[0038] Among them, the total reward totalReward is as follows:

[0039] totalReward=baseReward+negativeReward+downReward+successReward.

[0040] Specifically, the robotic arm includes a six-degree-of-freedom robotic arm, and the angle mapping relationship of each joint between the virtual experiment and the real environment is:

[0041] [θ′1, θ′2, θ′3, θ′4, θ′5, θ′6]=[θ1-180°, θ2+135°, θ3+90°, θ4-45°, θ5, θ6];

[0042] Among them, θ′1-θ′6 are the joint angles after transformation, and θ1-θ6 are the joint angles before transformation.

[0043] Specifically, the simulation platform adopts the CoppeliaSim platform.

[0044] Specifically, the tactile information includes the number and position data of the contact sensor points, that is, the contact area between the sole of the foot and the ground is represented by the number and position data of the collision sensor points;

[0045] When the pressure data of the sensing point exceeds the trigger threshold, the simulated ball corresponding to the sensing point is triggered, and it is determined that the sole of the foot is in contact with the ground.

[0046] As a second aspect, a tactile-based robot foot compliance system using the control method described above includes a simulated ground, a simulated foot robot, a plurality of simulated balls disposed on the soles of the simulated foot robot, an analysis and calculation module, a main control module, a posture measurement unit, and a tactile acquisition module disposed on the soles of the robot feet;

[0047] The tactile acquisition module collects tactile information from the sole of the actual foot robot and the ground, and obtains simulated tactile information by matching the tactile information with several simulated balls. The analysis and calculation module receives the simulated tactile information and analyzes and calculates it. The posture measurement unit uses the number of triggered sensor points to represent the simulated ground contact area of ​​the foot, which corresponds to the simulated foot robot posture information. The corresponding simulated posture information is mapped between the virtual experiment and the real environment through the main control module, and the posture of the end of the actual foot robot's mechanical arm is adjusted until the sole of the actual foot robot conforms to the ground.

[0048] Specifically, the posture measurement unit includes an inertial navigation module that integrates the posture and heading reference units, and represents the posture of the robot's feet by measuring the changes in the Euler angles of the foot of the foot robot.

[0049] The beneficial effects of the tactile-based robot foot ground compliance control method of the present invention are:

[0050] The present invention sets up several simulated small balls to simulate the sensor points on the soles of the actual foot robot to characterize the dynamic changes of tactile information in the movement of the robotic arm, and then obtains simulated control data through a control algorithm system based on deep reinforcement learning. The actual foot robot is dynamically adjusted to comply with the posture through a mapping relationship, thereby completing the compliance task of the foot robot with the ground. It autonomously adapts to the actual environment based on tactile information, and solves the problem in the existing technology that the robot lacks the ability to capture information such as contact pressure distribution with the ground, contact point position, and ground hardness, and relies too much on the robot's own sensation, resulting in insufficient adaptability and stability of the robot to some more special terrains. The present invention can perform balance control according to the robot's posture, greatly reduce the risk of instability of the robot, and ensure high stability and strong adaptability of the compliance. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0052] Figure 1 It is a flowchart of the steps of the control method according to an embodiment of the present invention.

[0053] Figure 2 This is the electrode arrangement diagram of the collection point of the tactile collection module according to the embodiment of the present invention.

[0054] Figure 3 It is a comparison diagram of some simulated ball triggering embodiments of the present invention.

[0055] Figure 4 FIG. 4 is a simplified schematic diagram of a simulated foot robot and a simulated ground surface according to an embodiment of the present invention.

[0056] Figure 5 This is a simplified diagram of rewards at different stages according to an embodiment of the present invention.

[0057] Figure 6 It is the three-dimensional relationship between s and the roll and pitch angles in the embodiment of the present invention.

[0058] Figure 7 It is a control flow chart of the control system of an embodiment of the present invention.

[0059] In the figure: 111, upper ground layer, 112, soft material layer, 113, lower reference layer, 12, simulated foot robot, 2, main control module, 3, tactile acquisition module, 4, analysis and calculation module. DETAILED DESCRIPTION

[0060] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0061] like Figure 1-Figure 3 The specific embodiment of the tactile-based robot foot ground compliance control method of the present invention includes the following steps:

[0062] S1. Build a foot motion simulation platform;

[0063] S2 is set on the simulation platform to simulate the foot robot 12 legs foot movement mechanism and plantar pressure sensing system;

[0064] S3. Collect and analyze tactile information during the actual foot robot plantar compliance process;

[0065] S4. Build a control algorithm system based on deep reinforcement learning to calculate and obtain simulated control data based on the collected and analyzed tactile information;

[0066] S5. Establish a mapping relationship between the virtual experiment and the real environment, that is, map the simulated control data to the actual foot robot and adjust the actual foot robot to conform to the posture.

[0067] The simulation platform includes a simulated ground, a simulated foot robot 12 and several simulated balls. The simulated balls are arranged and distributed on the soles of the simulated robot, corresponding to the sensor points in the sole pressure sensing system. The dynamic changes of tactile information in the actual movement of the robotic arm are represented according to the number of detected collision sensor points.

[0068] This embodiment takes a foot robot with a 6-DOF collaborative manipulator as an example. In this embodiment, the simulated ground adopts a double-layer simulation method, that is, the simulated ground is set as a double-layer simulation, including an upper ground layer 111, a soft material layer 112, and a lower reference layer 113 constructed from top to bottom. When the sensing point of the robot's foot touches the space between the upper ground layer 111 and the lower reference layer 113, it will be identified as a collision and simulated compression of the soft material layer 112. The actual foot robot foot has a layer of soft silicone as a buffering intermediate material, which gives it elastic deformation ability, so the robot foot is not completely rigid. An additional reference plane is added below the plane of the simulated ground. In this simulated ground, the newly introduced lower reference layer 113 and the area between the lower reference layer 113 and the upper ground layer 111 are simulated as the soft material layer 112. As shown in 4 and Figure 5 As shown, when the sensing point of the robot's foot contacts the space between the two layers, a collision is detected, and the soft material layer 112 is simulated to be compressed. The simulated ground plane is set to blue, and the reference plane is set to yellow. During the robot's simulated motion, a collision is detected when the sensing point on the actual robot's foot corresponding to the simulated ball contacts the two planes or the space between them.

[0069] The present invention combines double-layer simulation and small ball simulation, and can comprehensively consider the elastic deformation of the soft material on the sole of the actual foot robot and the contact pressure distribution between the sole of the robot and the ground, the contact point position, the ground hardness and other information. Through the control algorithm system based on deep reinforcement learning, it provides a controllable and realistic simulation platform for subsequent compliance control.

[0070] like Figure 2 As shown, in this embodiment, the diameter of the simulated ball is 3-7 mm. As a preferred embodiment, the diameter of the simulated ball is 5 mm. The plantar pressure sensing system includes a tactile acquisition module 3 for acquiring tactile information of the actual plantar of the foot robot and an analysis and calculation module 4 corresponding to the simulated ball. The tactile acquisition module 3 includes 128 sensor points arrayed on the plantar of the actual foot robot.

[0071] Specifically, the present invention creates a series of simulated balls with a diameter of 5 mm, and arranges them on the bottom of the robot's feet according to a specific arrangement pattern. These balls serve as sensing points and can move with the movement of the robot arm, thereby simulating the dynamic changes of tactile information during the movement of the robot arm. These simulated balls are designed to have only radius properties and no physical mass, so that they can perform effective collision detection without affecting the overall mass and motion characteristics of the robot's feet. The collision state is defined as a Boolean variable, with an initial state of 0 indicating no collision, and a state of 1 when a collision occurs. When these balls come into contact with other objects, the Boolean value changes from 0 to 1. In addition, the simulated balls that collide will change from their original color to red, as shown in the figure. Figure 3 As shown, the red balls not only mark the specific sensor points that collide with the ground, but their distribution also intuitively represents the distribution of the contact area. The program will capture and record these changes, including the sensor point number where the collision occurred and the total number of sensor points in contact with the environment.

[0072] The tactile acquisition module 3 in this embodiment uses a pressure sensor. The electrode layer of the pressure sensor is made of a flexible printed circuit (FPC), and polyimide (PI) is selected as the substrate. The copper plating process is selected to form electrodes on the PI substrate according to the design template. The double-layer electrode structure includes two groups of electrodes, one horizontal and one vertical. The two groups of electrodes cross to form capacitive sensing points, forming a sensing array. The leads of the vertical electrode group are led out from the heel part, and the leads of the horizontal electrode group are led out from the right side of the foot. The sensing points are distributed in 16 rows and 8 columns, and the connection method can be vertical or horizontal, see Figure 2 When in use, the pressure sensor is attached to the sole of the foot robot.

[0073] The tactile information in this embodiment includes the number and position data of contact sensor points. Specifically, the number and position data of contact sensor points represent the contact area between the sole of the foot and the ground. When the pressure data at a sensor point exceeds a trigger threshold, the sole of the foot is determined to be in contact with the ground. As an embodiment, the trigger threshold in this embodiment is set to 20. In the foot-ground contact scenario of the present invention, the reading of a sensor point is simplified to a binary switch value. That is, if the reading exceeds the set trigger threshold of 20, the area near that point is determined to be in contact with the ground. The present invention uses each sensor point as an indicator of whether the foot area it represents is in contact with the ground, rather than a precise pressure gauge. Specifically, in reality, a reading of a sensor point exceeding 20 corresponds to a value of 1 in simulation, indicating contact, and vice versa. In practical applications, the number of contact sensor points can represent the size of the foot-ground contact area. That is, the number of contact sensor points on the robot's sole is used to represent the contact area between the sole of the foot and the ground.

[0074] The simulation platform used in this embodiment is the CoppeliaSim platform. This open-source tool, based on physics engines such as Bullet, ODE, and Vortex, efficiently performs multi-scenario simulations. It provides a user-friendly visual interface and flexible scripting capabilities, making it relatively easy to control or simulate complex robotic systems. Highly modular and scalable, it supports a variety of robotic models, and its functionality can be extended through scripts or plug-ins.

[0075] The control algorithm system in this embodiment adopts a deep reinforcement learning algorithm based on the combination of value function and policy gradient.

[0076] The deep reinforcement learning algorithm consists of the following steps:

[0077] S401. Construct a reward and punishment model;

[0078] S402. Establish a state space; that is, the number of collisions between the sensor point and the upper and lower planes, the multiple joint angles of the manipulator, and the vertical coordinate Z of the center of the manipulator end are all included in the state space. The state space formula is:

[0079] S=[θ1,θ2,θ3,…θn,up_collision,bottom_collison,Z];

[0080] Among them, up_collision represents the number of collisions between the sensor point and the upper plane, bottom_collision represents the number of collisions between the sensor point and the lower plane, and θ1-θn represents multiple joint angles of the robot arm;

[0081] S403. Construct an action space; that is, define a multidimensional space that only includes the changes in the angles of multiple joints of the robotic arm. The corresponding formula for the action space A is:

[0082] A=[Δθ1, Δθ2, Δθ3,…, Δθn];

[0083] Among them, Δθ1, Δθ2, Δθ3, ..., Δθn are the angle changes of each joint of the robotic arm corresponding to each dimension in the next action cycle.

[0084] like Figure 4 and Figure 5 As shown, the reward and punishment model in this embodiment includes:

[0085] Base reward: When the sensor point is in contact only with the upper ground layer 111, a positive reward is assigned to the sensor point:

[0086] baseReward = up_collision;

[0087] Where baseReward is the base reward, and up_collision is the number of sensor points in contact with the upper ground.

[0088] Penalty mechanism: If a sensor point contacts the underlying reference surface, a negative reward is assigned to the sensor point:

[0089] negativeReward=-bottom_collison;

[0090] Where negativeReward is the penalty part, bottom_collision is the number of sensor points in contact with the reference plane;

[0091] Sparse rewards:

[0092] downReward=2.5×(pre_z-z);

[0093] Where z represents the vertical coordinate of the end of the manipulator, pre_z represents the vertical coordinate of the end of the manipulator at the beginning of each training cycle, and the unit is mm. If the foot moves in the negative direction of the z axis, a positive reward is given; on the contrary, if the foot moves in the positive direction of the z axis, it is considered a non-target behavior and a negative reward is given.

[0094] This approach introduces additional intermediate rewards to help the agent understand its state improvements. Specifically, rewards are given when an action brings the agent's new state closer to the goal than the previous one. The robot's foot interacts with the ground, so the reward function incorporates a determination of the foot's direction of movement. Specifically, as shown in Figure 5, the reward mechanism adjusts based on the foot's direction of movement relative to the z-axis, providing real-time feedback on the reward function. This prevents the agent from receiving very little or no reward in most states, only receiving significant rewards when specific goals or conditions are achieved.

[0095] SuccessReward reward mechanism: When the number of sensor points where the robot's foot detects a collision reaches 100% of the total number of sensor points, that is, the robot's foot is in full contact with the ground, the task is considered successfully completed and the SuccessReward value is set to +10. If this standard is not met, the reward value remains 0.

[0096] Among them, the total reward totalReward is as follows:

[0097] totalReward=baseReward+negativeReward+downReward+successReward.

[0098] In this embodiment, the robotic arm includes a six-degree-of-freedom robotic arm, and the angle mapping relationship between each joint of the virtual experiment and the real environment is:

[0099] [θ′1, θ′2, θ′3, θ′4, θ′5, θ′6]=[θ1-180°, θ2+135°, θ3+90°, θ4-45°, θ5, θ6];

[0100] Among them, θ′1-θ′6 are the joint angles after transformation, and θ1-θ6 are the joint angles before transformation.

[0101] The coordinate system in the simulation platform is set based on the algorithm's convenience, while the coordinate system in the real-world scenario is based on the actual position and orientation of the robotic arm. This difference in coordinate systems directly affects the calculation and representation of joint angles, leading to angle discrepancies between the simulation platform and reality. Before the real-world robotic arm angles are fed into the motion strategy module, the main program adjusts the angles of each joint by performing the aforementioned conversion based on the original angles to align the coordinate system differences between the simulation and real-world environments.

[0102] During the robot's motion, the main program continuously monitors tactile sensor data. If the data from any sensor point exceeds a safety threshold, the robot immediately stops and issues an alarm to prevent damage to the experimental equipment. The robot also stops if the running time exceeds a predetermined 50 time steps. Furthermore, the system sends notifications to the experimenter regarding successful mission completion and mission timeouts, ensuring smooth progress.

[0103] A tactile-based robot foot compliance control system using the above-mentioned control method includes a simulated ground, a simulated foot robot 12, several simulated balls arranged on the soles of the simulated foot robot 12, an analysis and calculation module 4, a main control module 2, a posture measurement unit, and a tactile acquisition module 3 arranged on the soles of the robot.

[0104] like Figure 7 As shown, the tactile acquisition module 3 collects tactile information from the actual foot robot's sole and the ground, and compares the tactile information with several simulated balls to obtain simulated tactile information. The analysis and calculation module 4 receives the simulated tactile information and analyzes and calculates it. The posture measurement unit uses the number of triggered sensor points to represent the simulated ground contact area of ​​the foot, corresponding to the posture information of the simulated foot robot 12. The corresponding simulated posture information is mapped between the virtual experiment and the real environment through the main control module 2, and the posture of the actual foot robot's mechanical arm is adjusted until the actual foot robot's sole conforms to the ground. The posture measurement unit includes an inertial navigation module (i.e., AHRS) that integrates the posture and heading reference unit. It represents the posture of the robot's foot by measuring the changes in the Euler angle of the foot.

[0105] In this embodiment, an AHRS is used to measure the robot's foot posture. A Python program communicates with the AHRS unit via a serial port, encapsulating this program into an interface. This interface is called in the main measurement program and used in conjunction with the pressure sensor measurement program. In the array measurement circuit code, count_over_20 = np.sum(sensingMatrix>20) is executed to count the number of points in the sensing matrix with a reading exceeding 20.

[0106] The AHRS unit is used to obtain the Euler angle data of the robot foot to characterize the posture of the robot foot. The Euler angle data includes the roll angle (roll angle), the pitch angle (pitch angle) and the contact area s. The three-dimensional relationship between s and the roll angle and pitch angle is as follows: Figure 6 As shown in the figure, it can be seen that when the roll angle and pitch angle both approach 0°, the area s reaches its peak. In view of the limitations of the sensor, the number of sensor points whose readings exceed the trigger threshold is used to represent the robot's foot-ground contact area. The larger the contact area between the foot and the ground, the higher its parallelism with the ground. At the same time, the two main conditions of compliant contact and parallelism are met, and the tactile pattern of the robot's foot-ground contact is obtained. During the data collection phase, the maximum reading of the sensor point reached 454. This paper sets a safety threshold slightly higher than this maximum value. Specifically, if the reading of any sensor point exceeds 511, it is considered to have occurred a potential unsafe contact. To ensure safety, once such a reading is detected, the robot's movement will be immediately interrupted. This measure prevents any situation that may cause damage or risk, ensuring the safety of the robot and the surrounding environment.

[0107] The main control module 2 in this embodiment includes a main program for storing the output command stream, a motion strategy unit for planning the robot arm motion strategy, and a robot arm controller for controlling the robot arm motion. In this embodiment, the robot arm is controlled through the joint_move mode, which has more direct control and faster execution response.

[0108] Analysis and Computation Module 4 performs computational analysis of the data and simultaneously operates the tactile sensors, collecting and converting tactile data into quantitative information in the simulation space. The specific process is as follows: First, the main program requests information about the six joint angles of the robotic arm. Based on this information and the robot's forward kinematics model, it calculates the spatial position of the robotic arm's end, while simultaneously acquiring tactile perception information from Analysis and Computation Module 4.

[0109] Considering that the tactile sensor is originally equipped with 128 sensing points, Figure 4 As shown, in this embodiment, a 3×5 sensor array is selected, which only includes 15 sensor points. To resolve this inconsistency, this paper adopts a method to convert the sum of the effective readings of the 128 sensor points (denoted as c) as follows.

[0110]

[0111] Where c′ is the effective reading after conversion. This conversion is intended to compress the larger sensor array data into effective information that conforms to the selected small array format to ensure consistency in data processing.

[0112] The criterion for mission success is defined as: all sensor points can detect collisions. However, according to the data in Section 3.2, when both the roll and pitch angles are close to 0°, that is, when the robot's sole is parallel to and in contact with the ground, the contact area typically exceeds 115, but does not reach the theoretical maximum of 128, meaning that readings are not guaranteed at all sensor points. Based on this, this section uses a contact area of ​​115 as the criterion for mission success. That is, if contact is detected at more than 115 sensor points, the mission is considered complete.

[0113] This state information is combined into a state st and fed into the motion strategy module. The module then outputs a 6-dimensional motion quantity at. The main program uses this information to calculate the robot's joint angles at the next moment and control the movement of the robot to the specified position. For safety and hardware stability reasons, the joint movement speed is set to 25% of the maximum speed. After each movement, the system pauses for 1.5 seconds before continuing.

[0114] The present invention sets up several simulated small balls to simulate the sensor points on the soles of the actual foot robot to characterize the dynamic changes of tactile information in the movement of the robotic arm, and then obtains simulated control data through a control algorithm system based on deep reinforcement learning. The actual foot robot is dynamically adjusted to comply with the posture through a mapping relationship, thereby completing the compliance task of the foot robot with the ground. It autonomously adapts to the actual environment based on tactile information, and solves the problem in the existing technology that the robot lacks the ability to capture information such as contact pressure distribution with the ground, contact point position, and ground hardness, and over-relies on the results of the robot's ground confidence detection, resulting in insufficient adaptability and stability of the robot to some more special terrains. The present invention can perform horizontal control according to the robot's posture, greatly reducing the risk of instability of the robot, and ensuring high stability and strong adaptability of the compliance.

[0115] It should be understood that the specific embodiments described above are only used to explain the present invention and are not intended to limit the present invention. Obvious changes or modifications derived from the spirit of the present invention are still within the scope of protection of the present invention.

Claims

1. A tactile-based robot foot ground compliance control method, characterized in that: The following steps are involved: S1. Build a foot motion simulation platform; S2. A mechanism for simulating foot movement of the robot (12) legs and a plantar pressure sensing system is provided on the simulation platform; S3. Collect and analyze tactile information during the actual compliance process of the foot robot; S4. Build a control algorithm system based on deep reinforcement learning to calculate and obtain simulated control data based on the collected and analyzed tactile information; S5. Establish a mapping relationship between the virtual experiment and the real environment, that is, map the simulated control data to the actual foot robot and adjust the actual foot robot to conform to the posture; The simulation platform includes a simulated ground, a simulated foot robot (12), and a plurality of simulated balls, wherein the simulated balls are arranged and distributed on the sole of the simulated robot, corresponding to the sensing points in the sole pressure sensing system, and the dynamic changes of the tactile information in the actual movement of the robot arm are represented according to the number of the detected collision sensing points; The simulated ground is configured as a double-layer simulation, comprising an upper ground layer (111), a soft material layer (112), and a lower reference layer (113) constructed sequentially from top to bottom; When the sensing point of the robot foot touches the space between the upper ground layer (111) and the lower reference layer (113), it is recognized as a collision and the compression of the soft material layer (112) is simulated; The control algorithm system adopts a deep reinforcement learning algorithm based on the combination of value function and policy gradient.

2. The tactile-based robot foot ground compliance control method according to claim 1, characterized in that: The diameter of the simulated ball is 3-7 mm; the plantar pressure sensing system comprises a tactile acquisition module (3) for acquiring tactile information of the plantar of an actual foot robot and an analysis and calculation module (4) corresponding to the simulated ball, wherein the tactile acquisition module (3) comprises 128 sensing points arrayed on the plantar of the actual foot robot.

3. The tactile-based robot foot ground compliance control method according to claim 1, characterized in that: The deep reinforcement learning algorithm includes the following steps: S401. Construct a reward and punishment model; S402. Establish a state space; that is, the number of collisions between the sensor point and the upper and lower planes, the multiple joint angles of the manipulator, and the vertical coordinate Z of the center of the manipulator end are incorporated into the state space. The state space formula is: S=[θ1,θ2,θ3,…θn,up_collision,bottom_collison,Z]; S=[θ1,θ2,θ3,θ4,θ5,θ6,up_collision,bottom_collison,z]; Among them, up_collision represents the number of collisions between the sensor point and the upper plane, bottom_collision represents the number of collisions between the sensor point and the lower plane, and θ1-θn represents multiple joint angles of the robot arm; S403. Construct an action space; that is, define a multidimensional space that only includes the changes in multiple joint angles of the robotic arm. The corresponding formula of the action space A is: A=[Δθ1, Δθ2, Δθ3,…, Δθn]; A=[Δθ1,Δθ2,Δθ3,Δθ4,Δθ5,Δθ6]; Among them, Δθ1, Δθ2, Δθ3, ..., Δθn are the angle changes of each joint of the robotic arm corresponding to each dimension in the next action cycle.

4. The tactile-based robot foot ground compliance control method according to claim 3, characterized in that: The reward and punishment model includes: Base reward: When the sensor point is in contact with only the upper ground layer (111), a positive reward is assigned to the sensor point: baseReward = up_collision; Where baseReward is the base reward, up_collision is the number of sensor points in contact with the upper ground; Penalty mechanism: If a sensor point contacts the underlying reference surface, a negative reward is assigned to the sensor point: negativeReward=-bottom_collison; Where negativeReward is the penalty part, bottom_collision is the number of sensor points in contact with the reference plane; Sparse rewards: downReward=2.5×(pre_z-z); Where z represents the vertical coordinate of the end of the manipulator, pre_z represents the vertical coordinate of the end of the manipulator at the beginning of each training cycle, and the unit is mm. If the foot moves in the negative direction of the z axis, a positive reward is given; on the contrary, if the foot moves in the positive direction of the z axis, it is considered a non-target behavior and a negative reward is given. SuccessReward reward mechanism: When the number of sensor points where the robot's foot detects a collision reaches 100% of the total number of sensor points, that is, the robot's foot is in full contact with the ground, the task is considered successfully completed and the SuccessReward value is set to +10. If this standard is not met, the reward value remains 0. Among them, the total reward totalReward is as follows: totalReward=baseReward+negativeReward+downReward+successReward.

5. The tactile-based robot foot ground compliance control method according to claim 1, characterized in that: The robotic arm includes a six-degree-of-freedom robotic arm, and the angle mapping relationship between each joint between the virtual experiment and the real environment is: [θ′1, θ′2, θ′3, θ′4, θ′5, θ′6]=[θ1-180°, θ2+135°, θ3+90°, θ4-45°, θ5, θ6]; Among them, θ′1-θ′6 are the joint angles after transformation, and θ1-θ6 are the joint angles before transformation.

6. The tactile-based robot foot ground compliance control method according to claim 1, characterized in that: The simulation platform adopts the CoppeliaSim platform.

7. The tactile-based robot foot ground compliance control method according to claim 1, characterized in that: The tactile information includes the number and position data of the contact sensor points, that is, the contact area between the sole of the foot and the ground is represented by the number and position data of the collision sensor points; When the pressure data of the sensing point exceeds the trigger threshold, the simulated ball corresponding to the sensing point is triggered, and it is determined that the sole of the foot is in contact with the ground.

8. A tactile-based robot foot compliance control system using the control method according to any one of claims 1 to 7, characterized in that: It comprises a simulated ground, a simulated foot robot (12), a plurality of simulated balls arranged on the soles of the simulated foot robot (12), an analysis and calculation module (4), a main control module (2), a posture measurement unit, and a tactile collection module (3) arranged on the soles of the robot; The tactile acquisition module (3) collects tactile information of the actual foot robot's sole and the ground, and obtains simulated tactile information by matching the tactile information with a number of simulated balls. The analysis and calculation module (4) receives the simulated tactile information and performs analysis and calculation on it. The posture measurement unit uses the number of triggered sensor points to represent the simulated ground contact area of ​​the foot, corresponding to the posture information of the simulated foot robot (12); the corresponding simulated posture information is mapped between the virtual experiment and the real environment through the main control module (2), and the posture of the end of the actual foot robot's mechanical arm is adjusted until the sole of the actual foot robot conforms to the ground.

9. The tactile-based robot foot compliance control system according to claim 8, characterized in that: The posture measurement unit includes an inertial navigation module that integrates the posture and heading reference units, and represents the posture of the robot's feet by measuring the changes in the Euler angles of the foot of the foot robot.

Images