A continuous assist control system and method for a hip exoskeleton

Abstract

The application relates to a continuous assistance control system and method of a hip exoskeleton, belonging to the technical field of robot control, and comprising a state generation subsystem, an SAC decision subsystem, a safe execution subsystem and a training data subsystem; LSTM time sequence state fusion modules input data collected by CNN visual modules and body sensation collection modules into an actor network after the data is fused, the actor network outputs a continuous action distribution, the continuous action distribution is mapped into torque or impedance action through an action mapping module, and is input into a low-level motor controller after being limited in amplitude and speed by a safety constraint layer, so that the left and right hip joint actuators execute assistance control instructions, and the training data subsystem collects assistance control data and performs offline correction on the SAC decision subsystem. The application repeatedly uses historical human-machine interaction data through an experience replay pool of an SAC optimal network, and is suitable for application scenarios with high exoskeleton data collection cost.

Description

Technical Field

[0001] This invention belongs to the field of robot control technology, specifically relating to a continuous assist control system for a hip exoskeleton and a continuous assist control method for a hip exoskeleton. Background Technology

[0002] The control objective of hip exoskeletons is to provide timely, smooth, energy-efficient, and safe assistive torque without interfering with the natural gait of the human body. Due to differences in wearer weight, walking speed, stride length, muscle strength, and terrain environment, traditional fixed-parameter controllers struggle to balance comfort, stability, and assistive effect across various scenarios. Reinforcement learning can learn control strategies through human-computer interaction data, but exoskeleton data acquisition is costly and carries a high risk of trial and error; therefore, algorithms with high sample efficiency and suitability for continuous motion spaces are required.

[0003] Existing methods for controlling hip exoskeletons include phase control based on finite state machines, trajectory tracking control based on PID or impedance models, intention-triggered control based on electromyography or plantar pressure, and policy control based on reinforcement learning algorithms such as DQN and PPO. While these methods can provide assistance in specific scenarios, they still have limitations in situations involving continuous terrain changes, individual wearer differences, and visual prediction.

[0004] Existing finite state machine (FSM) control schemes typically switch between different assist curves based on gait phase, such as applying a preset torque during hip flexion or extension. Impedance control schemes output spring-damped torque based on joint deviations. In reinforcement learning schemes, DQN typically outputs discrete actions, such as increasing assist, decreasing assist, or switching modes, and is mainly suitable for discrete action spaces. If used for exoskeleton control, the torque or mode usually needs to be discretized. PPO can output continuous actions, but the training data is difficult to reuse, and the cost of parameter tuning on actual machines is high.

[0005] Furthermore, existing reinforcement learning exoskeleton control schemes often directly output the policy network to the motor driver, lacking an independent safety constraint layer. During the exploration phase, this may result in excessive torque, rapid torque changes, or movements that are not comfortable for the human body. Without safety limits, rate constraints, and anomaly fallback mechanisms, it is difficult to meet the engineering requirements of wearable human-computer interaction devices. Especially for medical assistive and wearable devices, repeatedly collecting real human-computer interaction data is costly and requires strict safety constraints, which existing technical solutions cannot meet. Summary of the Invention

[0006] The purpose of this invention is to address the problems existing in the prior art by providing a continuous assist control system and method for a hip exoskeleton. The continuous assist control method based on SAC enables the exoskeleton to receive the comprehensive spatiotemporal state output by LSTM, generate continuous actions by combining maximum entropy separation strategy reinforcement learning, and convert the strategy actions into executable hip joint motor control commands through an independent safety constraint layer, anomaly backoff logic, and a low-level motor controller.

[0007] The technical solution is as follows: A continuous assist control system for a hip exoskeleton includes a state generation subsystem, a SAC (Soft Actor-Critic) decision subsystem, a safe execution subsystem, and a training data subsystem. The state generation subsystem includes a CNN vision module, a proprioception acquisition module, and an LSTM (Soft Actor-Critic) temporal state fusion module. The SAC decision subsystem includes an actor network, a dual critic network, a target network, and an entropy-temperature module. The safe execution subsystem includes a motion mapping module, a safety constraint layer, an anomaly backoff module, a low-level motor controller, and left and right hip joint actuators. The training data subsystem includes a reward calculation module, an experience playback pool, and an offline / background update module. The LSTM temporal state fusion module fuses the data collected by the CNN vision module and the proprioception acquisition module and inputs it into the actor network. The actor network outputs a continuous motion distribution, which is mapped into torque or impedance motion by the motion mapping module. After amplitude and rate limiting by the safety constraint layer, the data is input into the low-level motor controller to control the left and right hip joint actuators to execute assist control commands. The training data subsystem collects assist control data and performs offline corrections to the SAC decision subsystem.

[0008] Furthermore, in the state generation subsystem, the CNN vision module receives images from a forward-facing camera or a depth camera and outputs terrain probability, obstacle information, and visual confidence as inputs to the LSTM and reward weight adjustment; the proprioception acquisition module acquires IMU, joint angles, joint velocities, motor current, temperature, plantar pressure, or safety status as real-time feedback inputs for continuous assist control; and the LSTM temporal state fusion module fuses visual features and proprioceptive sequences to output a comprehensive spatiotemporal state, gait phase, and uncertainty.

[0009] Furthermore, in the SAC decision subsystem, the actor network obtains normalized actions by compressing the mean and variance of the continuous action distribution of the state output using a hyperbolic tangent function; a dual critic network, including two value networks Q1 and Q2, is used to estimate the value of state actions and suppress overestimation by taking the smaller value, while maintaining the target critic network; the entropy temperature module automatically adjusts the entropy temperature coefficient alpha to maintain a balance between exploration and stability in the strategy.

[0010] Furthermore, in the safety execution subsystem, the motion mapping module maps the normalized motion output by the actor network into the left and right hip joint torque or impedance parameters that the exoskeleton can execute; the safety constraint layer limits the torque amplitude, torque change rate, impedance parameter range, and bilateral coordination constraints of the original motion execution, and immediately switches to zero assistance or conservative impedance control through the abnormal backoff module when encountering the preset safety event termination condition; the low-level execution module converts the safety motion into motor control commands and executes left and right hip joint assistance in a high-frequency closed loop.

[0011] Furthermore, the conditions for terminating a safety incident include: emergency stop, communication timeout, excessive temperature, joint overtravel, and failure of continuous safety projection.

[0012] Furthermore, in the training data subsystem, the reward calculation module generates instant rewards based on gait symmetry, assist timing, joint power, motion stability, energy consumption, torque smoothness, and safety events. The weight coefficients of each reward item are adaptively adjusted in real time based on the terrain probability vector output by the CNN vision module. The experience replay pool records the state, action, reward, next state, safety correction amount, and anomaly flags for offline training, live fine-tuning, and fault tracing.

[0013] Furthermore, the weighting of the reward items is adjusted as follows: when the probability of detecting a slope is highest, the stability reward weight and energy consumption penalty weight are automatically increased; when walking on flat ground, the gait symmetry reward weight and assist timing reward weight are automatically increased; when stairs or obstacles are detected, the safety event penalty weight is automatically increased; the reward weighting adjustment formula is: ,in For the first Each reward item is in Adaptive weights at any given time; As the benchmark weight; Terrain category For reward items The amount of weight adjustment; Terrain categories output by the CNN vision module The real-time probability.

[0014] A continuous assist control method for a hip exoskeleton, using the above system and performing the following steps: Step 100: Complete the training of the SAC (Soft Actor-Critic) network framework to obtain the optimal SAC network; Step 200: Deploy the SAC optimization network into the controller of the hip exoskeleton. The optimization network consists of an actor network, a motion mapping module, and a safety constraint layer. Step 300: The controller of the hip exoskeleton acquires the comprehensive spatiotemporal state of the LSTM output in each control cycle and reads the current joint angle, velocity, motor current, temperature and safety status. Step 400: Input the integrated spatiotemporal state into the SAC optimization network, output the action distribution parameters, and apply the formula: Sampling yields normalized continuous actions, where For normalized continuous actions; and These are the mean function and standard deviation function output by the actor network, respectively; The synthesized spatiotemporal state output by the LSTM; This is sampling noise; Used to compress actions to a normalized range; Step 500: Map the normalized motion to the exoskeleton's executable range to form the original motion; Step 600: The safety constraint layer performs amplitude and rate limiting. Amplitude and rate limiting use a control barrier function to project the original motion to the nearest point in the set of safe motions, rather than simply truncating it. The projection formula for the control barrier function is: ;in, The original motion output for SAC actors via network. For safety procedures after projection, For the first The control barrier function corresponding to each safety constraint, Here are the CBF constraint convergence coefficients. For safety constraints on quantity; Step 700: The low-level execution module performs motor control of the hip exoskeleton; Step 800: The left and right hip joint actuators execute the power assist control command according to the motor control; Step 900: Collect the status, human feedback or agent indicators for the next moment, and calculate the immediate reward. If a security anomaly occurs, proceed to step 1000; if there is no security anomaly, proceed to step 1100. Step 1000: Switch between conservative assist or zero assist mode, then return to step 600; Step 1100: Write to the experience recycling pool for offline policy updates to complete continuous assist control of the hip exoskeleton.

[0015] Furthermore, the training of the SAC network framework in step 100 includes the following steps: Step 110: Construct the SAC network framework by combining the actor network, dual critic network, target critic network, entropy temperature module, and experience replay pool, and randomize the pre-training strategy; Step 120: Collect interaction data and randomize wearer mass, walking speed, terrain slope, friction coefficient, sensor noise and execution latency to improve policy transfer capability; Step 130: Initialize the experience replay pool using existing exoskeleton experimental data or data collected by a conservative controller, and sample training batches. The formula for the experience replay data structure is as follows: .in, For empirical samples; and These represent the current state and the next state, respectively. For action; For instant rewards; This is the end marker; For safety signs; Step 140: Use the formula: Calculate the critic's target value and update the dual critic network and the target critic network, where the next action is sampled from the target policy. For the critic's target value; For instant rewards; Discount factor; and For two target commentator networks; It is the entropy temperature coefficient; For actor strategy distribution; Step 150: Minimize the critic's mean squared error loss, maximize the actor's target value and entropy term, and automatically update the actor network and entropy temperature coefficient based on the target entropy. The formula is: ,in, For the critics' loss, Aim for acting; Estimated by the commentator network; The target value; and These represent the entropy temperature coefficient and the strategy distribution, respectively. Step 160: Verify whether the actual machine indicators of the action output meet the requirements through the safety constraint layer. If the requirements are not met, proceed to step 170; if the requirements are met, proceed to step 180. Step 170: Reduce exploration variance and adjust reward weights, then return to step 150; Step 180: Complete the training of the SAC network framework to obtain the optimal SAC network.

[0016] Furthermore, the safety constraint layer in step 600 also includes bilateral coordinated constraints: ,in The preset maximum allowable difference in torque between the two sides, and the gait phase coupling constraint, that is, when the left leg is in the support phase, the right assist torque It should not exceed the preset proportion of the left leg's supporting torque, and vice versa, to ensure that the assistance from both sides is mechanically coordinated.

[0017] Beneficial effects: 1) This invention reuses historical human-computer interaction data through an experience replay pool via the SAC optimization network, which is suitable for application scenarios with high exoskeleton data acquisition costs. The strategy directly outputs continuous torque or impedance parameters, avoiding mechanical shock and wearing discomfort caused by discrete mode switching.

[0018] 2) The maximum entropy objective allows the strategy to maintain appropriate randomness during the training phase, which is beneficial for adapting to different wearers and terrain changes.

[0019] 3) The safety constraint layer is independent of the policy network and imposes hard constraints on torque, speed, joint range and abnormal states, reducing the risk of actual machine operation.

[0020] 4) The reward weights are adaptively adjusted based on the terrain probability, enabling the SAC strategy to automatically focus on the most critical optimization objectives in different scenarios. It prioritizes stability on slopes, gait symmetry on flat ground, and safety in front of obstacles, avoiding the scenario preference limitations of fixed-weight schemes. Attached Figure Description Figure 1 This is a flowchart illustrating the continuous assist control method for the hip exoskeleton of the present invention. Figure 2 A schematic diagram of the training process for the SAC network framework. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit it. Terms such as "upper," "lower," "front," "rear," "left," "right," "bottom," "inner," and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the purpose of describing the invention. They do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention.

[0022] A continuous assist control system for a hip exoskeleton includes a state generation subsystem, a SAC (Soft Actor-Critic) decision subsystem, a safe execution subsystem, and a training data subsystem. The state generation subsystem includes a CNN vision module, a proprioception acquisition module, and an LSTM (Soft Actor-Critic) temporal state fusion module. The SAC decision subsystem includes an actor network, a dual critic network, a target network, and an entropy-temperature module. The safe execution subsystem includes a motion mapping module, a safety constraint layer, an anomaly backoff module, a low-level motor controller, and left and right hip joint actuators. The training data subsystem includes a reward calculation module, an experience playback pool, and an offline / background update module. The LSTM temporal state fusion module fuses the data collected by the CNN vision module and the proprioception acquisition module and inputs it into the actor network. The actor network outputs a continuous motion distribution, which is mapped into torque or impedance motion by the motion mapping module. After amplitude and rate limiting by the safety constraint layer, the data is input into the low-level motor controller to control the left and right hip joint actuators to execute assist control commands. The training data subsystem collects assist control data and performs offline corrections to the SAC decision subsystem.

[0023] In the state generation subsystem, the CNN vision module receives images from a forward-facing camera or depth camera and outputs terrain probability, obstacle information, and visual confidence as inputs to the LSTM and reward weight adjustment. The proprioception acquisition module acquires IMU, joint angles, joint velocities, motor current, temperature, plantar pressure, or safety status as real-time feedback inputs for continuous assist control. The LSTM temporal state fusion module fuses visual features and proprioceptive sequences to output a comprehensive spatiotemporal state, gait phase, and uncertainty.

[0024] In the SAC decision subsystem, the actor network outputs the mean and variance of the continuous action distribution based on the state, and then compresses it using a hyperbolic tangent function to obtain normalized actions; a dual critic network, including two value networks, Q1 and Q2, is used to estimate the value of state actions and suppress overestimation by taking the smaller value, and to maintain the target critic network; the entropy temperature module automatically adjusts the entropy temperature coefficient alpha to maintain a balance between exploration and stability in the strategy.

[0025] In the safety execution subsystem, the motion mapping module maps the normalized motion output from the actor network into executable left and right hip joint torques or impedance parameters for the exoskeleton. The safety constraint layer limits the torque amplitude, torque change rate, impedance parameter range, and bilateral coordination constraints on the original motion execution. Upon encountering preset safety event termination conditions, it immediately switches to zero-assistance or conservative impedance control via an abnormal rollback module. The lower-level execution module converts the safety motion into motor control commands and executes left and right hip joint assistance in a high-frequency closed-loop manner. Safety event termination conditions include: emergency stop, communication timeout, excessive temperature, joint boundary violation, and continuous safety projection failure.

[0026] In the training data subsystem, the reward calculation module generates real-time rewards based on gait symmetry, assist timing, joint power, motion stability, energy consumption, torque smoothness, and safety events. The weight coefficients of each reward item are adaptively adjusted in real time based on the terrain probability vector output by the CNN vision module. The experience replay pool records the state, action, reward, next state, safety correction, and anomaly flags for offline training, live fine-tuning, and fault tracing. The weight coefficients of the reward items are adjusted as follows: when the probability of detecting a slope is highest, the stability reward weight and energy consumption penalty weight are automatically increased; when walking on flat ground, the gait symmetry reward weight and assist timing reward weight are automatically increased; when stairs or obstacles are detected, the safety event penalty weight is automatically increased. The reward weight adjustment formula is: ,in For the first Each reward item is in Adaptive weights at any given time; As the benchmark weight; Terrain category For reward items The amount of weight adjustment; Terrain categories output by the CNN vision module The real-time probability.

[0027] like Figure 1 The present invention discloses a continuous assist control method for a hip exoskeleton, which uses the above system and performs the following steps: Step 100: Complete the training of the SAC network framework to obtain the optimal SAC network; Step 200: Deploy the SAC optimization network into the controller of the hip exoskeleton. The optimization network consists of an actor network, a motion mapping module, and a safety constraint layer. Step 300: The controller of the hip exoskeleton acquires the comprehensive spatiotemporal state of the LSTM output in each control cycle and reads the current joint angle, velocity, motor current, temperature and safety status. Step 400: Input the integrated spatiotemporal state into the SAC optimization network, output the action distribution parameters, and apply the formula: Sampling yields normalized continuous actions, where For normalized continuous actions; and These are the mean function and standard deviation function output by the actor network, respectively; The synthesized spatiotemporal state output by the LSTM; This is sampling noise; Used to compress actions to a normalized range; Step 500: Map the normalized motion to the exoskeleton's executable range to form the original motion; Step 600: The safety constraint layer performs amplitude and rate limiting. Amplitude and rate limiting use a control barrier function to project the original motion to the nearest point in the set of safe motions, rather than simply truncating it. The projection formula for the control barrier function is: ;in, The original motion output for SAC actors via network. For safety procedures after projection, For the first The control barrier function corresponding to each safety constraint, Here are the CBF constraint convergence coefficients. The number of safety constraints; the safety constraint layer also includes bilateral coordination constraints: ,in The preset maximum allowable difference in torque between the two sides, and the gait phase coupling constraint, that is, when the left leg is in the support phase, the right assist torque It should not exceed the preset proportion of the left leg's supporting torque, and vice versa, to ensure that the assistance from both sides is mechanically coordinated. Step 700: The low-level execution module performs motor control of the hip exoskeleton; Step 800: The left and right hip joint actuators execute the power assist control command according to the motor control; Step 900: Collect the status, human feedback or agent indicators for the next moment, and calculate the immediate reward. If a security anomaly occurs, proceed to step 1000; if there is no security anomaly, proceed to step 1100. Step 1000: Switch between conservative assist or zero assist mode, then return to step 600; Step 1100: Write to the experience replay pool for offline strategy updates to complete continuous assist control of the hip exoskeleton.

[0028] like Figure 2 As shown, training the SAC network framework includes the following steps: Step 110: Construct the SAC network framework by combining the actor network, dual critic network, target critic network, entropy temperature module, and experience replay pool, and randomize the pre-training strategy; Step 120: Collect interaction data and randomize wearer mass, walking speed, terrain slope, friction coefficient, sensor noise and execution latency to improve policy transfer capability; Step 130: Initialize the experience replay pool using existing exoskeleton experimental data or data collected by a conservative controller, and sample training batches. The formula for the experience replay data structure is as follows: .in, For empirical samples; and These represent the current state and the next state, respectively. For action; For instant rewards; This is the end marker; For safety signs; Step 140: Use the formula: Calculate the critic's target value and update the dual critic network and the target critic network, where the next action is sampled from the target policy. For the critic's target value; For instant rewards; Discount factor; and For two target commentator networks; It is the entropy temperature coefficient; For actor strategy distribution; Step 150: Minimize the critic's mean squared error loss, maximize the actor's target value and entropy term, and automatically update the actor network and entropy temperature coefficient based on the target entropy. The formula is: ,in, For the critics' loss, Aim for acting; Estimated by the commentator network; The target value; and These represent the entropy temperature coefficient and the strategy distribution, respectively. Step 160: Verify whether the actual machine indicators of the action output meet the requirements through the safety constraint layer. If the requirements are not met, proceed to step 170; if the requirements are met, proceed to step 180. Step 170: Reduce exploration variance and adjust reward weights, then return to step 150; Step 180: Complete the training of the SAC network framework to obtain the optimal SAC network.

[0029] Example: A collimated drive motor is installed in each hip joint. The bottom-level controller executes a torque closed loop at 500 Hz, and the SAC strategy outputs the action at 50 Hz. The LSTM module outputs a 128-dimensional integrated spatiotemporal state, along with gait phase and uncertainty. The actor network consists of two fully connected layers, each with 256 neurons, outputting a 6-dimensional continuous action: expected torque of the left hip, expected torque of the right hip, stiffness gain of the left hip, stiffness gain of the right hip, damping gain of the left hip, and damping gain of the right hip. The action is compressed using a hyperbolic tangent function and mapped to a safe range. The safety constraint layer uses a control obstacle function projection instead of traditional amplitude limiting, and simultaneously constrains the upper limit of torque, the upper limit of torque change rate, the joint angle limit, and the upper limit of motor temperature. When the original SAC action does not meet the safety conditions, the following formula is used: To solve for the safety correction action, CBF projection parameters Choose between 1.0 and 2.0 to balance safety margin and response speed.

[0030] In addition, the system enforces bilateral coordination constraints and gait phase coupling constraints. When a communication timeout is detected, the system automatically switches the action to zero assist or conservative impedance. The reward function consists of five parts, and the weight coefficients of each part are adaptively adjusted in real time according to the CNN terrain classification output. The weight adjustment is smooth and the weight change rate between adjacent control cycles does not exceed 0.05, avoiding policy oscillation caused by sudden weight changes.

[0031] During the training phase, pre-training is first performed in a simulation environment using domain randomization, followed by offline replay training using real data collected by a conservative controller, and finally fine-tuning is performed on a live machine with a small exploration variance. During the deployment phase, the critic network and the target network do not participate in real-time control; only the actor network is retained. Actions are output using mean or low-variance sampling and always pass through a safety constraint layer.

[0032] Verification has shown that this embodiment can provide continuous and smooth assistance in scenarios such as going up and down slopes, accelerating on flat ground, and crossing low obstacles.

[0033] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the principles and spirit of the present invention should be included within the protection scope of the present invention.

Claims

1. A continuous assist control system for a hip exoskeleton, characterized in that: It includes a state generation subsystem, a SAC decision subsystem, a safe execution subsystem, and a training data subsystem. The state generation subsystem includes a CNN vision module, a proprioception acquisition module, and an LSTM temporal state fusion module. The SAC decision subsystem includes an actor network, a dual critic network, an object network, and an entropy temperature module. The safe execution subsystem includes an action mapping module, a safety constraint layer, an anomaly backoff module, a low-level motor controller, and left and right hip joint actuators. The training data subsystem includes a reward calculation module, an experience replay pool, and an offline / background update module. The LSTM temporal state fusion module fuses the data collected by the CNN vision module and the proprioception acquisition module and inputs it into the actor network. The actor network outputs a continuous motion distribution, which is mapped into torque or impedance motion by the motion mapping module. After the safety constraint layer performs amplitude and rate limiting, it is input into the low-level motor controller to control the left and right hip joint actuators to execute assist control commands. The training data subsystem collects assist control data and performs offline correction on the SAC decision subsystem.

2. The continuous assist control system for the hip exoskeleton as described in claim 1, characterized in that: In the state generation subsystem, the CNN vision module receives images from a forward-facing camera or a depth camera and outputs terrain probability, obstacle information, and visual confidence as inputs to LSTM and reward weight adjustment. The proprioception acquisition module collects data from the IMU, joint angle, joint speed, motor current, temperature, plantar pressure, or safety status as real-time feedback input for continuous assist control. The LSTM temporal state fusion module fuses visual features and proprioceptive sequences to output a comprehensive spatiotemporal state, gait phase, and uncertainty.

3. The continuous assist control system for the hip exoskeleton as described in claim 1, characterized in that: In the SAC decision subsystem, the actor network outputs the mean and variance of the continuous action distribution based on the state, and then compresses it using a hyperbolic tangent function to obtain normalized actions; a dual critic network, including two value networks, Q1 and Q2, is used to estimate the value of state actions and suppress overestimation by taking the smaller value, and to maintain the target critic network; the entropy temperature module automatically adjusts the entropy temperature coefficient alpha to maintain a balance between exploration and stability in the strategy.

4. The continuous assist control system for the hip exoskeleton as described in claim 1, characterized in that: The motion mapping module in the safety execution subsystem maps the normalized motion output by the actor network into left and right hip joint torque or impedance parameters that the exoskeleton can execute. The safety constraint layer limits the torque amplitude, torque change rate, impedance parameter range, and bilateral coordination constraints of the original motion. When a preset safety event termination condition is encountered, the abnormal backoff module immediately switches to zero assistance or conservative impedance control. The low-level execution module converts the safety motion into motor control commands and executes left and right hip joint assistance in a high-frequency closed loop.

5. The continuous assist control system for the hip exoskeleton as described in claim 4, characterized in that: The aforementioned safety event termination conditions include: emergency stop, communication timeout, excessive temperature, joint overtravel, and failure of continuous safety projection.

6. The continuous assist control system for the hip exoskeleton as described in claim 1, characterized in that: In the training data subsystem, the reward calculation module generates instant rewards based on gait symmetry, assist timing, joint power, motion stability, energy consumption, torque smoothness, and safety events. The weight coefficients of each reward item are adaptively adjusted in real time according to the terrain probability vector output by the CNN vision module. The experience replay pool records the state, action, reward, next state, safety correction amount, and anomaly flag, which can be used for offline training, live fine-tuning, and fault tracing.

7. The continuous assist control system for the hip exoskeleton as described in claim 6, characterized in that: The weighting adjustment method for the reward items is as follows: when the probability of detecting a slope is highest, the stability reward weight and energy consumption penalty weight are automatically increased; when walking on flat ground, the gait symmetry reward weight and assist timing reward weight are automatically increased; when stairs or obstacles are detected, the safety event penalty weight is automatically increased; the reward weight adjustment formula is: ,in For the first Each reward item is in Adaptive weights at any given time; As the benchmark weight; Terrain category For reward items The amount of weight adjustment; Terrain categories output by the CNN vision module The real-time probability.

8. A method for continuous assist control of a hip exoskeleton, characterized in that: Using the continuous assist control system of the hip exoskeleton as described in any one of claims 1-7 and performing the following steps: Step 100: Complete the training of the SAC network framework to obtain the optimal SAC network; Step 200: Deploy the SAC optimization network into the controller of the hip exoskeleton. The optimization network consists of an actor network, a motion mapping module, and a safety constraint layer. Step 300: The controller of the hip exoskeleton acquires the comprehensive spatiotemporal state of the LSTM output in each control cycle and reads the current joint angle, velocity, motor current, temperature and safety status. Step 400: Input the integrated spatiotemporal state into the SAC optimization network, output the action distribution parameters, and apply the formula: Sampling yields normalized continuous actions, where For normalized continuous actions; and These are the mean function and standard deviation function output by the actor network, respectively; The synthesized spatiotemporal state output by the LSTM; This is sampling noise; Used to compress actions to a normalized range; Step 500: Map the normalized motion to the exoskeleton's executable range to form the original motion; Step 600: The safety constraint layer performs amplitude and rate limiting. Amplitude and rate limiting use a control barrier function to project the original motion to the nearest point in the set of safe motions, rather than simply truncating it. The projection formula for the control barrier function is: ;in, The original motion output for SAC actors via network. For safety procedures after projection, For the first The control barrier function corresponding to each safety constraint, Here are the CBF constraint convergence coefficients. For safety constraints on quantity; Step 700: The low-level execution module performs motor control of the hip exoskeleton; Step 800: The left and right hip joint actuators execute the power assist control command according to the motor control; Step 900: Collect the status, human feedback or agent indicators for the next moment, and calculate the immediate reward. If a security anomaly occurs, proceed to step 1000; if there is no security anomaly, proceed to step 1100. Step 1000: Switch between conservative assist or zero assist mode, then return to step 600; Step 1100: Write to the experience replay pool for offline strategy updates to complete continuous assist control of the hip exoskeleton.

9. The continuous assist control method for the hip joint exoskeleton as described in claim 8, characterized in that: The training of the SAC network framework in step 100 includes the following steps: Step 110: Construct the SAC network framework by combining the actor network, dual critic network, target critic network, entropy temperature module, and experience replay pool, and randomize the pre-training strategy; Step 120: Collect interaction data and randomize wearer mass, walking speed, terrain slope, friction coefficient, sensor noise and execution latency to improve policy transfer capability; Step 130: Initialize the experience replay pool using existing exoskeleton experimental data or data collected by a conservative controller, and sample training batches. The formula for the experience replay data structure is as follows: .in, For empirical samples; and These represent the current state and the next state, respectively. For action; For instant rewards; This is a termination marker; For safety signs; Step 140: Use the formula: Calculate the critic's target value and update the dual critic network and the target critic network, where the next action is sampled from the target policy. For the critic's target value; For instant rewards; Discount factor; and For two target commentator networks; It is the entropy temperature coefficient; For actor strategy distribution; Step 150: Minimize the critic's mean squared error loss, maximize the actor's target value and entropy term, and automatically update the actor network and entropy temperature coefficient based on the target entropy. The formula is: ,in, A loss for critics, Aim for acting; Estimated by the commentator network; The target value; and These represent the entropy temperature coefficient and the strategy distribution, respectively. Step 160: Verify whether the actual machine indicators of the action output meet the requirements through the safety constraint layer. If the requirements are not met, proceed to step 170; if the requirements are met, proceed to step 180. Step 170: Reduce exploration variance and adjust reward weights, then return to step 150; Step 180: Complete the training of the SAC network framework to obtain the optimal SAC network.

10. The continuous assist control method for the hip exoskeleton as described in claim 8, characterized in that: The security constraint layer in step 600 includes both sides of coordinated constraints: ,in The preset maximum allowable difference in torque between the two sides, and the gait phase coupling constraint, that is, when the left leg is in the support phase, the right assist torque. It should not exceed the preset proportion of the left leg's supporting torque, and vice versa, to ensure that the assistance from both sides is mechanically coordinated.

Images