A robot joint torque self-adjusting control system based on deep learning

Abstract

The application relates to the technical field of robot control, in particular to a robot joint torque self-adjusting control system based on deep learning, which comprises a robot main body, an environment perception and coordinate construction module, a deep learning torque modeling module, a path planning and torque pre-adjusting module, a real-time feedback correction module and a multi-joint collaborative optimization module; data interaction is realized among the modules through a CAN-Ethernet hybrid bus.In the scheme, through the collaborative collection of a laser radar, a depth camera, a joint force sensor and an IMU, combined with the conversion of a world coordinate system, a main body coordinate system and a joint local coordinate system, the integration of environment features and motion parameters in a unified framework is realized.This design avoids torque deviation caused by non-uniform coordinates or data fragmentation, makes torque control more in line with the actual environment and the robot state, improves the overall accuracy of torque control from the basic level, and adapts to the precise demand of torque calculation in a complex scene.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, and in particular to a robot joint torque self-adjustment control system based on deep learning. Background Technology

[0002] In today's rapidly developing robotics technology, joint torque control is a core component for achieving precise movement and safe operation in robots, and its performance directly determines the robot's applicability in complex scenarios. Whether it's assembly operations in the industrial field, human-robot interaction in the service sector, or autonomous movement in special scenarios, robots are required to adjust joint torques in real time according to environmental changes to balance motion accuracy, stability, and energy consumption. As robot applications become increasingly diversified, the types of obstacles in the environment increase, and the complexity of motion paths rises, placing higher demands on the dynamic response capability of joint torque adjustment, multi-joint coordination, and environmental adaptability.

[0003] However, existing robot joint torque control systems have significant shortcomings. On the one hand, the correlation between environmental perception and robot body coordinates is low, and multi-source sensor data is often fragmented, making it difficult to form a unified coordinate framework. This leads to deviations in the baseline of torque calculation, which in turn affects the accuracy of torque adjustment. On the other hand, torque modeling often relies on a single algorithm, failing to simultaneously consider environmental spatial characteristics and motion temporal patterns, resulting in insufficient rationality of initial torque calculations. Furthermore, real-time feedback correction mechanisms are lagging, and multi-joint collaborative optimization capabilities are weak. This not only easily leads to control lag effects but also makes it difficult to respond quickly to sudden obstacles, failing to meet the requirements for efficient and safe robot movement in complex scenarios. Therefore, we propose a deep learning-based self-adjusting control system for robot joint torque. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a deep learning-based robot joint torque self-adjustment control system, thereby solving the technical problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A robot joint torque self-adjustment control system based on deep learning includes a robot body, an environment perception and coordinate construction module, a deep learning torque modeling module, a path planning and torque pre-adjustment module, a real-time feedback correction module, and a multi-joint collaborative optimization module.

[0007] The main body of the robot is used to perform tasks and move;

[0008] The environmental perception and coordinate construction module is used to collect environmental information and robot state data, and to establish a spatial coordinate system;

[0009] The deep learning torque modeling module is configured to calculate initial joint torques based on environmental features and coordinate parameters through a deep learning model;

[0010] The path planning and torque pre-adjustment module is configured to generate a pre-planned path and pre-adjust joint torques according to the path;

[0011] The real-time feedback correction module is configured to collect real-time motion data, calibrate torques, and respond to sudden obstacles;

[0012] The multi-joint collaborative optimization module is configured to achieve global coordination and time sequence iterative control of multi-joint torques;

[0013] The modules interact with each other through a CAN-Ethernet hybrid bus.

[0014] In one possible implementation, the robot body includes a left arm execution area, a right arm execution area, a left leg support area, a right leg support area, and a body driving area;

[0015] The left arm execution area and the right arm execution area are located on both sides above the torso, used for operation and task execution, and the end is configured with a force-sensitive clamping device;

[0016] The left leg support area and the right leg support area are located on both sides below the torso, used for support and movement, and driven by a harmonic reducer and a servo motor;

[0017] The body driving area is located at the center of the torso, integrated with a main controller, a power management unit, and a high-speed communication interface.

[0018] In one possible implementation, the environment perception and coordinate construction module collects environmental information and robot state data through multi-source sensors, and establishes a three-level coordinate system of a world coordinate system, a body coordinate system, and a joint local coordinate system; the multi-source sensors include a laser radar, a depth camera, a joint force sensor, and an inertial measurement unit (IMU); the environment perception and coordinate construction module outputs an environmental feature set.

[0019] In one possible implementation, the deep learning torque modeling module uses a CNN-LSTM hybrid deep learning model to calculate an initial torque vector based on an input feature vector fused from environmental features, robot parameters, and coordinate features, and to impose constraints on the output torque;

[0020] The deep learning torque modeling module further calculates initial target angles and initial motion velocities of each joint by solving the inverse kinematics problem through the Newton-Raphson method.

[0021] In one possible implementation, the path planning and torque pre-adjustment module uses an improved The algorithm generates a pre-planned path and discretizes the expected position and expected joint angle according to the control cycle;

[0022] The path planning and torque pre-adjustment module adjusts the torque by combining PID control and nonlinear compensation;

[0023] The path planning and torque pre-adjustment module also verifies the physical feasibility of the pre-adjusted torque to ensure that the angular acceleration does not exceed the maximum bearing value of the joint.

[0024] In a possible implementation, the real-time feedback correction module collects actual angles, actual torques, and real-time environment data, and calculates angle deviations and torque deviations;

[0025] The module uses an incremental learning support vector machine (ISVM) algorithm for dynamic calibration, and outputs the calibrated torque of the joint.

[0026] The real-time feedback correction module also has an emergency response mechanism. When a sudden obstacle is detected and the shortest distance is <0.3m, the movement speed is adjusted or a braking torque is applied.

[0027] In a possible implementation, the multi-joint cooperative optimization module constructs a multi-objective optimization function ,

[0028] ;

[0029] wherein, is the cooperative objective function, is the set of associated joint pairs, is the cooperative coefficient of the joint pair, and is the torque set of the jth joint and the ith joint, is the smoothness objective function, is the weight coefficient, is the calibrated torque of the last cycle; The multi-joint cooperative optimization module uses an improved particle swarm optimization (PSO) algorithm to solve the optimal torque set;

[0030] The multi-joint cooperative optimization module performs time sequence iteration according to the control cycle until the robot completes the task.

[0031] The multi-joint cooperative optimization module performs time sequence iteration according to the control cycle until the robot completes the task.

[0032] Compared with the prior art, the beneficial effects are:

[0033] ​​1. In this scheme, by constructing a multi-source sensor fusion environment perception system and a three-level coordinate system, the problem of accurately associating environmental information with the robot's state is effectively solved. Through the cooperative collection of laser radar, depth camera, joint force sensor and IMU, combined with the conversion of world coordinate system, body coordinate system and joint local coordinate system, the integration of environmental features and motion parameters in a unified framework is realized. This design avoids the torque deviation caused by non-unified coordinates or data fragmentation, allowing torque control to better match the actual environment and robot state, improving the overall accuracy of torque control from a basic level, and adapting to the precise torque calculation requirements in complex scenarios;

[0034] 2. In this scheme, a CNN-LSTM hybrid deep learning model is used for torque modeling, and a path pre-adjustment strategy is combined to significantly improve the rationality of initial torque calculation and the forward-looking nature of torque adjustment. The CNN part can effectively extract spatial correlation features in the environment, and the LSTM part can learn the time sequence law of torque changes with motion state, and the combination of the two realizes nonlinear accurate mapping of environment and torque. At the same time, based on the improved algorithm, path planning and PID combined with nonlinear compensation pre-adjustment can adapt to path changes and detour requirements in advance, reducing the hysteresis effect of the control system, and making torque adjustment more in line with the dynamic needs of the robot's actual motion process;

[0035] 3. In this scheme, through the real-time feedback correction mechanism of incremental learning support vector machine and the multi-joint collaborative optimization strategy of improved particle swarm optimization, the safety and overall performance of robot motion are greatly improved. In the real-time feedback stage, the dynamic calibration algorithm does not need to retrain historical samples, and can quickly adapt to deviation changes. The emergency response mechanism designed for sudden obstacles can adjust the motion parameters in time to ensure safety. The multi-joint collaborative optimization takes motion stability and energy balance as the goal, realizes the global coordination of multi-joint torque, and makes the robot maintain stability in complex motion, efficiently utilize energy, and flexibly respond to sudden conditions, further expanding the application space of robots in complex scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0036] The above description is only a summary of the technical scheme of the present application. In order to more clearly understand the technical means of the present application and can be implemented according to the content of the description, the following will be described in detail with the preferred embodiments of the present application and with the help of the accompanying drawings.

[0037] Fig. 1 The system framework diagram of the self-adjusting control system of the present application;

[0038] Fig. 2 The robot body schematic diagram of the self-adjusting control system of the present application;

[0039] Fig. 3A schematic diagram of a self-adjusting control system for controlling robot movement according to the present application.

[0040] Legend: 100, left arm execution area; 200, right arm execution area; 300, left leg support area; 400, right leg support area; 500, main body driving area. DETAILED DESCRIPTION

[0041] The preferred embodiments of the present application will be described in detail by referring to the attached drawings, however the present application can be realized in various different forms, therefore the present application is not limited to the embodiments described below, and in order to more clearly describe the present application, components not connected to the invention will be omitted from the drawings;

[0042] The technical solutions in the embodiments of the present application are to solve the problems in the above background art, and are generally as follows:

[0043] Embodiments

[0044] Please refer to Figs. 1 to 3 The present embodiment takes a six-degree-of-freedom serial joint robot as an application carrier, and introduces a robot joint torque self-adjusting control system based on deep learning, which includes a robot main body, an environment perception and coordinate construction module, a deep learning torque modeling module, a path planning and torque pre-adjustment module, a real-time feedback correction module, and a multi-joint collaborative optimization module.

[0045] The robot main body includes a left arm execution area 100, a right arm execution area 200, a left leg support area 300, a right leg support area 400, and a main body driving area 500, wherein the left arm execution area 100 and the right arm execution area 200 are located on both sides above the torso, representing the left and right arms of the robot, responsible for operation and task execution, and are configured with force-sensitive clamping devices at the end; the left leg support area 300 and the right leg support area 400 are located on both sides below the torso, representing the left and right legs of the robot, responsible for support and movement, and are driven by harmonic reducers and servo motors; the main body driving area 500 is located at the center of the torso, integrating a main controller, a power management unit, and a high-speed communication interface. Six joints are provided in each area to assist the robot main body in movement.

[0046] The environment perception and coordinate construction module is used to collect environment information and robot state data, and establish a spatial coordinate system; the deep learning torque modeling module calculates the initial joint torque based on environmental features and coordinate parameters through deep learning; the path planning and torque pre-adjustment module is used to generate a pre-planned path and pre-adjust the torque according to the path; the real-time feedback correction module is used to collect real-time motion data, calibrate the torque, and respond to sudden obstacles; and the multi-joint collaborative optimization module is used to realize global coordination and time sequence iterative control of multi-joint torque.

[0047] The data interaction between the modules is realized through a CAN-Ethernet hybrid bus, and the modules cooperate with each other to form a running process for the movement of the robot body, which is as follows:

[0048] S1: Environment perception and coordinate system construction

[0049] The environment perception and coordinate construction module collects the surrounding environment information of the robot and the motion state data of the robot through multiple source sensors, and synchronously establishes a three-axis coordinate system with the robot body as the reference. Specifically, the origin is set as the geometric center of the main driving area, that is, the installation position of the inertial measurement unit (IMU); The Y-axis points to the front along the central axis of the robot body, and is consistent with Fig. 3 the forward direction of the robot, and is used to represent the displacement change in the front-back direction; The Z-axis (not marked in the figure) is perpendicular to the Y-axis and located on the horizontal plane, and points to the right side of the robot. This axis is a special axis for left-right translation, and when an obstacle with a height > 0.3 m or a width > 0.8 m is encountered, the displacement adjustment in the Y-axis direction is used to realize detouring; The X-axis is perpendicular to the YZ plane and points upward, and is consistent with the Z-axis, the Y-axis, and the X-axis form a right-handed coordinate system, which is used to represent the position change in the vertical direction.

[0050] The hardware assembled in this module includes a laser radar, a depth camera, a joint force sensor, and an inertial measurement unit (IMU). The layout and data processing logic of each hardware are as follows: the laser radar is horizontally installed at the top front end of the main driving area, and the scanning plane is 0.6 m away from the ground. It performs three-dimensional scanning on the front 180° range at a frequency of 10 Hz, and generates raw point cloud data. The raw point cloud data is preprocessed. First, the ground and low noise points are removed through straight-through filtering (points with X-axis coordinates > 0.1 m are retained), and then the discrete point cloud is aggregated into an obstacle entity through a clustering algorithm. The minimum circumscribed rectangle parameters of the obstacle are extracted to form an obstacle coordinate set . The depth camera is installed directly below the laser radar, with the lens facing forward. It synchronously collects the environment RGB image and depth information. Through a semantic segmentation algorithm, the obstacle attributes (such as rigid walls, flexible fabrics, and mobile devices) are identified, and the three-dimensional contour features of the target object are extracted to determine the center coordinates of the target object in the image coordinate system. Combined with the camera intrinsic parameters and the depth value , the initial spatial coordinates are converted. The joint force sensor is built-in at the output end of each joint, with a total of 24, which can collect the torque value In the robot power-on initialization phase (static state), record the static torque of each joint As the torque calculation reference. IMU is integrated in the geometric center of the main body driving area, collects the real-time attitude angle (roll angle , pitch angle , yaw angle ) and motion acceleration of the robot, and removes high-frequency vibration interference through complementary filtering algorithm, and outputs smooth attitude and acceleration data.

[0051] To realize the coordinate association between the environment space and the robot body, the world coordinate system is established synchronously, whose origin is the fixed reference point in the environment, the axis coincides with the axis in the initial state of the robot, the axis coincides with the axis in the initial state, and the axis is perpendicular to the ground and upward. Through the identification and positioning of the fixed reference point by the laser radar, combined with the yaw angle output by the IMU, the conversion matrix from the robot body coordinate system to the world coordinate system is calculated:

[0052] ;

[0053] Among them, is the coordinate of the origin of the body coordinate system in the world coordinate system, the value directly reflects the translation distance of the robot in the left and right directions, and is the core parameter to realize the detour function. In addition, a local coordinate system is established at the rotation center of each joint, and the conversion matrix of the local coordinate system and the body coordinate system is calculated through the D-H parameters of the robot , forming a three-level coordinate system of "world coordinate system-body coordinate system-joint local coordinate system", ensuring that all motion parameters are calculated in a unified framework.

[0054] After this step, the environment perception and coordinate construction module outputs the environment feature set ( is the obstacle attribute label), the coordinate conversion parameter set and the initial torque reference . The above data is transmitted to the deep learning torque modeling module through the high-speed bus.

[0055] S2: Deep learning torque modeling and generating initial torque

[0056] The deep learning torque modeling module takes the environmental features and coordinate parameters output by S1 as input, combines the physical properties of the robot joints, and calculates the initial target torque of each joint through a deep learning model.

[0057] First, feature fusion processing is performed to integrate environmental features, robot parameters, and coordinate features into a model input vector. The environmental features include the shortest distance in the axis direction of the obstacle in the subject coordinate system The offset in the axis direction (positive value indicates that the obstacle is on the right side of the robot, negative value indicates on the left side), the coordinates of the target object relative to the subject coordinate system (converted through the transformation matrix ); the robot parameters include the moment of inertia , friction coefficient , link mass , and joint zero error ; the coordinate features mainly extract the motion trend in the axis direction, when the absolute value of the obstacle is less than 0.5m, the weight of is increased to 1.5 times that of other features. The above features are normalized (mapped to the interval) to construct an input feature vector with a dimension of 1x64 .

[0058] The deep learning model used is a CNN-LSTM hybrid structure. This model extracts spatial features and learns temporal associations to achieve nonlinear mapping of the environment-torque. The CNN part of the model includes 3 convolutional layers and 2 pooling layers. The convolutional layer uses a 3x3 convolutional kernel to extract spatial correlation information in the environmental features. The pooling layer uses a 2x2 maximum pooling window, and the output dimension is a spatial feature vector of 1x32 . The LSTM part includes 2 layers of recurrent units, each with 128 hidden nodes, learning the temporal rules of torque changes with motion state. After fusing the spatial feature vector and the historical torque features of the joints, the output dimension is an initial torque vector of 1x24 , corresponding to the 24 joints of the robot. To ensure torque output safety, constraints are imposed on the model output , where the initial target torque of the joint, , and are determined by the rated parameters of the joint motor, which are the minimum allowed torque of the joint and the maximum allowed torque of the joint, respectively.​​ The maximum allowable torque of each joint, the output value beyond the range is truncated to the threshold boundary.

[0059] The initial joint angle is calculated by inverse kinematics, with the coordinates of the target object relative to the coordinate system of the main body The forward kinematics equation is established based on the D-H parameters of the robot, and the inverse kinematics problem is solved by Newton-Raphson method to obtain the initial target angle of each joint . At the same time, set the initial motion speed according to the task requirements , wherein, , , are the initial motion speeds of the robot along the axis, axis, axis of the coordinate system of the main body, respectively Dynamic adjustment according to the When >0, <0 (left translation detour); when <0, >0 (right translation detour), the absolute value of the translation speed is inversely proportional to the absolute value of .

[0060] This step finally generates the initial control instruction set , which is transmitted to the path planning and torque pre-adjustment module through the high-speed bus.

[0061] S3: Path planning and torque pre-adjustment

[0062] The path planning and torque pre-adjustment module is based on the initial control instruction output by S2, and the path is predicted for the next 3 cycles with a control cycle of 10ms, and the joint torque is pre-adjusted according to the prediction result, so as to reduce the lag effect of the control system on the robot, and adapt to the detour requirement in the axis direction.

[0063] The path planning adopts an optimization algorithm based on improved , taking the current position of the robot (determined by the coordinates of the origin of the coordinate system of the main body in the world coordinate system) as the starting point, and the coordinates of the target object in the world coordinate system as the end point, to generate multiple candidate paths in the world coordinate system. When sampling the path, the feasibility in the axis direction is considered: if the distance of the obstacle in the axis direction is <1m and the height is >0.3m (cannot be crossed), then in the The axis direction deviates from the original straight path to generate a left or right detour path, and the detour offset is determined according to the width of the obstacle; if the obstacle can be crossed, the original straight path is kept unchanged. The candidate paths are evaluated, and the evaluation indexes include the total length of the path, The maximum axis direction offset, the path curvature change rate, and the comprehensive optimal path are selected by weighted summation as the pre-planned path The pre-planned path is discretized according to the control period to obtain the expected position of the next 3 periods The expected joint angle corresponding to each expected position is calculated by kinematics forward solution , wherein, is the expected joint angle of the current period.

[0064] Based on the pre-planned path, torque pre-regulation is performed, and a strategy combining PID control and nonlinear compensation is adopted. The PID controller takes the difference between the expected joint angle of the current period and the actual joint of the previous period as input, calculates the torque basic adjustment amount , and the torque basic adjustment amount calculation formula of the first joint is:

[0065] ;

[0066] wherein, , , is the PID parameter of the first joint, is the proportional coefficient of the PID controller of the first joint, is the integral coefficient of the PID controller of the first joint, is the differential coefficient of the PID controller of the first joint, and all are determined by online fuzzy self-tuning, is the difference between the expected joint angle of the first joint in the current period and the actual joint in the previous period. The nonlinear compensation part is aimed at the changes of ground friction and load disturbance when detouring left and right, and adopts a single-hidden-layer neural network to output the compensation amount , which takes the joint angular velocity , the distance of the obstacle in the axis direction and the current torque error as input, and outputs a compensation amount vector with a dimension of 1x24. The final pre-regulation torque of the first joint is , wherein is the first ​The compensation amount of the joint is output by a single hidden layer neural network. The pre-adjustment torque is physically verified, and the actual output torque corresponding to the angular acceleration is calculated by the joint dynamics model , to ensure that it does not exceed the maximum bearing value of the joint. The pre-adjustment torque is substituted into the dynamics equation , wherein, is the moment of inertia of the first joint, is the damping coefficient of the first joint, is the angular velocity of the first joint, is the gravity torque of the first joint, and is obtained by solving. If the calculated exceeds the maximum angular acceleration threshold of the joint (uniformly set to 5 rad / s 2 ), the torque is recalculated according to the maximum angular acceleration and the is corrected.

[0067] The pre-planned path segment , the expected joint angle and the pre-adjustment torque of the current period are output in this step and transmitted to the real-time feedback correction module through the high-speed bus.

[0068] S4: Real-time feedback correction and emergency response to obstacles

[0069] The real-time feedback correction module collects current motion state data through sensors, compares it with the expected parameters output by S3, calculates the deviation and dynamically calibrates the torque, and simultaneously performs emergency response to sudden obstacles to ensure the safety of the robot motion.

[0070] The sensor data collection and processing process is as follows: the joint encoder collects the actual angles of each joint at a frequency of 1 kHz , and the angular velocity is calculated by dividing the angle difference between two adjacent sampling periods by the sampling period; the joint force sensor synchronously collects the actual torque of each joint ; the laser radar and depth camera update the environment data in real time, focusing on monitoring the position change of the obstacle, and calculating the real-time distance between the obstacle and the robot . All data are transmitted to the controller after time stamp alignment, and the controller performs effectiveness verification on the data, and eliminates data that exceeds the sensor range or has continuous 3 sampling period abnormal jumps. After data processing is completed, two types of core deviations are calculated: angle deviation and torque deviation Differential deviation thresholds are set based on the different functions of the joints: the angular deviation threshold for the operated joint (arm joint) is 0.3°, and the torque deviation threshold is 0.2°. The threshold for angular deviation of the drive joint (leg joint) is 0.8°, and the threshold for torque deviation is 0.5°. Maintain the pre-adjustment torque when all joint deviations are within the threshold range. If the deviation of at least one joint exceeds the threshold, the dynamic calibration process is initiated.

[0071] Dynamic calibration employs the incremental learning support vector machine (ISVM) algorithm. This algorithm does not require retraining on all historical samples; it dynamically updates the decision boundary only by adding new real-time samples. The input feature vector of the algorithm is... ,in, For the first Obstructions near the joint The real-time distance along the axis is output as the first... Joint torque calibration During model initialization, an initial ISVM is trained using 1000 sets of offline-collected multi-scene bias-calibration samples. In real-time operation, the bias data from the current period is combined with the optimal calibration values ​​to form new samples, dynamically updating the model parameters. This is to adapt... In the scenario of axis bypass, when the robot is in Axis translation state When >0, dynamically increase the calibration sensitivity of the leg joint and the main drive joint. After calibration, the [number]th [unit / section]... The torque of the joint is ,in, For the first The pre-adjustment torque of the joint will Substituting the angular acceleration into the dynamic equations to verify the first... Joint angular acceleration ≤5rad / s 2 .

[0072] When the lidar detects a sudden obstacle, and the obstacle is at its shortest distance from the robot... When the distance is less than 0.3m, the emergency response mechanism is triggered. The specific operation is as follows: First, calculate the direction of movement of the sudden obstacle. With speed magnitude Predict the location of the obstacle within the next two control cycles. If the predicted location intersects with the robot's future path, adjust immediately. Component of motion in the axial direction: when the obstacle moves towards When moving in the positive direction of the axis, increase the robot's position along the main coordinate system. Initial velocity of the shaft When the obstruction is towards When moving in the negative direction of the axis, the robot's movement along the body coordinate system is reduced Initial movement speed of the axis At the same time, the movement speed of the axis is reduced to 50% of the original speed; if the movement of the axis in both directions is blocked, control all joints to generate braking torque , , = 1.2 to stop the robot within a safe distance. During the emergency response process, the environment feature set is updated in real time and the pre-planned path .

[0073] This step outputs the calibrated torque and the expected joint angle of the next cycle , and transmits it to the multi-joint collaborative optimization module through the high-speed bus.

[0074] S5: Multi-joint collaborative optimization and time sequence iteration

[0075] The multi-joint collaborative optimization module is based on the calibrated torque and the expected joint angle output by S4, and realizes the global coordination of multi-joint torque through a multi-objective optimization algorithm, taking into account the motion smoothness and energy balance, while performing time sequence iteration according to the control cycle until the robot completes the predetermined task.

[0076] Construct a multi-objective optimization function , with the coordination of joint torque and motion smoothness as the core optimization target, the expression is:

[0077] ;

[0078] Among them, is the coordination target function, is the set of associated joint pairs, is the coordination coefficient of the joint pair, and are the torque sets of the th joint and the th joint;

[0079] Smoothness target function, is the calibrated torque of the last cycle; is the weight coefficient. In the axis detour scene = 0.7, in the straight line motion scene = 0.3.

[0080] The optimization process needs to meet three types of constraint conditions: centroid stability constraint, the centroid height of the robot is ≥ 0.3 m, and the center of mass is in The offset in the axis direction ≤ 0.05 m; joint torque constraints, each joint torque needs to meet ; energy constraints, the total energy consumption of the robot does not exceed 20% of the current remaining power.

[0081] An improved particle swarm optimization (PSO) algorithm is used to solve the above multi-objective optimization problem, which balances the global search and local convergence ability through adaptive adjustment of the inertia weight. In the initialization stage of the algorithm, 50 particles are generated, each particle representing a set of torque solutions , so the first dimension represents the torque dimension of the first joint; in the iteration process, each particle updates the speed and position according to the individual optimal position and the global optimal position .

[0082] The speed update formula is:

[0083] ;

[0084] where, is the speed of the first dimension in the first iteration, is the inertia weight, which balances the global search and local convergence ability of the algorithm, is the speed of the first dimension in the first iteration, , are acceleration constants, is the individual learning factor, which controls the tendency of the particle to the individual optimal position, is the social learning factor, which controls the tendency of the particle to the global optimal position, , are random numbers, usually uniformly generated in , is the optimal value of the first dimension torque of the particle in the historical iteration, is the optimal value of the first dimension torque of all particles in the historical iteration, is the position of the first dimension of the particle in the first iteration;

[0085] The position update formula is , where, is the position of the first dimension of the particle in the first iteration.

[0086] After 20 iterations, output the torque set corresponding to the globally optimal position. This signal is then converted into a PWM control signal to drive the actuators at each joint.

[0087] With a control cycle of 10ms, processes S3 to S5 are executed repeatedly to form a timing control chain. At the end of each cycle, the actual position of the robot body in the world coordinate system is calculated. , For the first At the end of each control cycle, the robot's main body origin... The actual position of the target object is calculated. European distance , For the first The Euclidean distance between the actual position of the robot body and the position of the target object at the end of each cycle. When the deviation is ≤0.05m, the task is considered complete, and the robot is controlled to perform the end effector operation; otherwise, the control process continues to the next cycle. In the time-series iteration, if the angle deviation or torque deviation exceeds the threshold for 5 consecutive cycles, the online fine-tuning mechanism of the deep learning torque modeling module is triggered, thereby ensuring that the model always maintains high control accuracy.

[0088] Finally, it should be noted that the above embodiments are merely examples for clearly illustrating the present invention and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A robot joint torque self-adjustment control system based on deep learning, characterized in that, It includes the robot body, environmental perception and coordinate construction module, deep learning torque modeling module, path planning and torque pre-adjustment module, real-time feedback correction module, and multi-joint collaborative optimization module; The main body of the robot is used to perform tasks and move; The environmental perception and coordinate construction module is used to collect environmental information and robot state data, and to establish a spatial coordinate system; The deep learning torque modeling module is used to calculate the initial joint torque based on environmental features and coordinate parameters using a deep learning model. The path planning and torque pre-adjustment module is used to generate a pre-planned path and pre-adjust the joint torque according to the path; The real-time feedback correction module is used to collect real-time motion data, calibrate torque, and respond to sudden obstacles. The multi-joint collaborative optimization module is used to realize the global coordination and time-series iterative control of multi-joint torques; The modules communicate with each other via a CAN-Ethernet hybrid bus. The path planning and torque pre-adjustment module uses an improved A* algorithm to generate a pre-planned path, and discretizes it according to the control cycle to obtain the desired position and desired joint angle. The path planning and torque pre-adjustment module performs torque pre-adjustment through a strategy combining PID control and nonlinear compensation. The path planning and torque pre-adjustment module also performs physical feasibility verification on the pre-adjustment torque to ensure that the angular acceleration does not exceed the maximum bearing value of the joint. The real-time feedback correction module collects actual angle, actual torque and real-time environmental data, and calculates angle deviation and torque deviation. The real-time feedback correction module uses the incremental learning support vector machine (ISVM) algorithm for dynamic calibration, and outputs the calibrated result. The torque of the joint; The real-time feedback correction module is also equipped with an emergency response mechanism. When a sudden obstacle is detected and the shortest distance is <0.3m, the movement speed is adjusted or a braking torque is applied. The multi-joint collaborative optimization module constructs a multi-objective optimization function. , ； in, For the cooperative objective function, For the set of associated joint pairs, denoted as the coordination coefficient of the joint pair. and For the first The joint and the first The set of torques at each joint Let the objective function be stationarity. These are the weighting coefficients. This is the calibration torque for the previous cycle; The multi-joint collaborative optimization module uses an improved particle swarm optimization (PSO) algorithm to solve for the optimal torque set. The multi-joint collaborative optimization module iterates sequentially according to the control cycle until the robot completes the task.

2. The robot joint torque self-adjustment control system based on deep learning as described in claim 1, characterized in that, The robot body includes a left arm execution area (100), a right arm execution area (200), a left leg support area (300), a right leg support area (400), and a main body drive area (500). The left arm execution area (100) and the right arm execution area (200) are located on both sides above the torso and are used for operation and task execution. The ends are equipped with force-sensitive clamping devices. The left leg support area (300) and the right leg support area (400) are located on both sides below the torso for support and movement, and are driven by a harmonic reducer and a servo motor. The main drive area (500) is located at the center of the torso and integrates a main controller, a power management unit and a high-speed communication interface.

3. The robot joint torque self-adjustment control system based on deep learning as described in claim 1, characterized in that, The environment perception and coordinate construction module collects environmental information and robot state data through multi-source sensors and establishes a three-level coordinate system consisting of a world coordinate system, a body coordinate system, and a joint local coordinate system. The multi-source sensors include lidar, a depth camera, joint force sensors, and an inertial measurement unit (IMU). The environment perception and coordinate construction module outputs an environmental feature set.

4. The robot joint torque self-adjustment control system based on deep learning as described in claim 1, characterized in that, The deep learning torque modeling module adopts a CNN-LSTM hybrid deep learning model to calculate the initial torque vector using an input feature vector that fuses environmental features, robot parameters, and coordinate features, and applies constraints to the output torque. The deep learning torque modeling module also calculates the initial target angle and initial motion velocity of each joint by solving the inverse kinematics problem using the Newton-Raphson method.

Images