A robot agent evaluation system for lower extremity exoskeleton and a corresponding gait strategy training method of the agent

Abstract

The present application relates to a kind of robot agent evaluation system for lower extremity exoskeleton and the corresponding gait strategy training method of agent, belong to rehabilitation medical auxiliary equipment technical field.The device includes humanoid robot, exoskeleton, auxiliary device and detection element, exoskeleton is connected with humanoid robot by auxiliary device, first auxiliary structure is provided with recess, second auxiliary structure is provided with protrusion, recess is connected with humanoid robot, protrusion is detachably connected with humanoid robot;Detection element is installed in the sole of humanoid robot.Utilize gait strategy training method to adjust the gait of humanoid robot, so that it can imitate the walking posture of different healthy people or different hemiplegia patients etc.By wearing exoskeleton on the device and deploying different walking strategies, the relevant data of the change of each joint angle, torque, power, overall energy consumption and gait symmetry of humanoid robot under different walking conditions is obtained, which helps exoskeleton developer to better test and optimize exoskeleton product.

Description

Technical Field

[0001] This invention relates to a robotic agent evaluation system for lower limb exoskeletons and a corresponding gait strategy training method, belonging to the field of rehabilitation medical assistive device technology. Background Technology

[0002] Lower limb exoskeleton robots are wearable devices that assist users in walking or enhance human function. The quality of their control strategies directly determines the effectiveness of walking assistance and the user experience. Currently, the performance evaluation of exoskeleton walking strategies mainly relies on human trials. Commonly used evaluation indicators include metabolic consumption and electromyography (EMG) signals. However, metabolic consumption testing, which measures oxygen consumption, requires subjects to wear expensive breathing masks for extended periods of walking. The testing process is cumbersome and costly, and the results are easily affected by the subject's physiological state, psychological factors, and individual differences, resulting in poor repeatability and consistency. EMG signals, on the other hand, suffer from high signal noise and difficulty in accurately corresponding joint torque and power, leading to insufficient accuracy in evaluation.

[0003] To overcome the limitations of human testing, the industry has begun exploring the use of sensor-equipped robotic dummies for exoskeleton testing. Dummies can provide standardized mechanical feedback, avoiding the problems associated with individual differences. However, existing dummy solutions have significant shortcomings: First, dummies are typically passive structures, requiring external suspension systems to simulate gravity, and cannot achieve active walking. Therefore, they are mainly suitable for evaluating fully articulated exoskeletons, and are not effective for evaluating exoskeletons that only assist single joints such as the hip or knee. Second, dummy systems are complex, bulky, and limited in testing scenarios, lacking flexibility.

[0004] With the maturation of general-purpose humanoid robot technology, these robots possess biomimetic structures and active movement capabilities similar to humans, enabling them to simulate realistic human gait. This characteristic makes humanoid robots ideal standardized "test subjects." However, the current lumbar structure of humanoid robots is not suitable for wearing exoskeletons, and there are no detailed, standardized, quantifiable, and repeatable evaluation methods for assistive walking strategies. Furthermore, there is a lack of detailed methods for evaluating these strategies, as well as methods to enable humanoid robots to mimic human gait and provide more accurate test data to help exoskeleton developers better test and understand their exoskeleton-assisted walking or rehabilitation strategies.

[0005] Therefore, there is an urgent need for an auxiliary device that connects humanoid robots and exoskeletons, as well as an evaluation method for assistive walking strategies based on general humanoid robots, so as to achieve quantitative evaluation of exoskeleton assistive walking strategies in a low-cost, high-precision, and repeatable manner. Summary of the Invention

[0006] To address the aforementioned issues, this invention provides a robot agent evaluation system for lower limb exoskeletons and a corresponding gait strategy training method, which enables quantitative evaluation of exoskeleton assistive walking strategies in a low-cost, high-precision, and repeatable manner.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a robotic agent evaluation system for a lower limb exoskeleton, comprising: A humanoid robot exoskeleton detection device, the detection device comprising: a humanoid robot, an exoskeleton, an auxiliary device, and a detection element, wherein the exoskeleton is connected to the humanoid robot via the auxiliary device, and the detection element is installed on the sole of the humanoid robot's foot; The robotic proxy evaluation system for the lower limb exoskeleton includes the following steps: Step 1: System Setup; First, install the detection element on the sole of the humanoid robot's foot, install the exoskeleton to be evaluated on the humanoid robot through the auxiliary device, and ensure that the joint axis of the exoskeleton is roughly aligned with the corresponding lower limb joint axis of the humanoid robot. Establish a communication connection between the detection element, the exoskeleton controller, and the main control system of the humanoid robot, and develop a data acquisition system to collect data from the three in real time. Step 2: Gait strategy training; train different walking strategies for the humanoid robot according to the assessment content; if the test is to test the assistive walking strategy of normal people, the walking trajectory training model of normal people is used as the benchmark strategy; if the test content is to help hemiplegic patients recover, improve gait symmetry or assist walking, the maximum joint torque of unilateral and bilateral joints is limited by modifying the reinforcement learning reward function, and the gait data of hemiplegic patients in the database is used for imitation learning. Step 3: Baseline data acquisition; control the humanoid robot to walk along a preset standard test path at a set speed curve; during this process, keep the exoskeleton in zero-torque mode; through the humanoid robot's built-in sensors, collect and record the joint angles θ of each lower limb joint in real time throughout the walking process. baseline Joint angular velocity ω baseline Joint torque τ baseline And data from the built-in IMU that changes over time; Step 4: Assistance Strategy Testing; Under the exact same test environment, path, and set speed curve, activate the exoskeleton's assistance strategy to be tested, controlling the humanoid robot to walk again with the same gait parameters as in Step 3. Similarly, collect and record the joint angles θ of each lower limb joint under this exoskeleton assistance strategy in real time. baseline Joint angular velocity ω baseline Joint torque τ baselineAnd data from the built-in IMU that changes over time; Step 5: Performance Index Calculation; Based on the data collected in Steps 3 and 4, calculate multi-dimensional quantitative evaluation indicators for the exoskeleton-assisted walking strategy, including assistive efficiency. Gait symmetry and auxiliary performance indicators; all calculations are performed per unit mass, i.e., by dividing the corresponding physical quantity by the total mass of the humanoid robot body and exoskeleton. This ensures the objectivity and universality of the evaluation criteria.

[0008] In one embodiment of the present invention, if the exoskeleton is used to evaluate its mobility assistance capabilities, the evaluation index is assist efficiency. The aforementioned assist efficiency The calculation formula is as follows:

[0009] in, The energy consumed to complete a specific task without wearing an exoskeleton. The energy consumed to perform the same task while wearing an exoskeleton. and It can be calculated directly from humanoid robot data.

[0010] In one embodiment of the present invention, if the exoskeleton is used to evaluate the rehabilitation effect on hemiplegic patients, then the core performance evaluation is gait symmetry. The gait symmetry The calculation formula is as follows:

[0011] in, For gait symmetry, The average support period time of the left leg under multiple gait states is the average of the entire time from the heel touching the ground to the toes leaving the ground in each gait cycle. The support phase time of the right leg is the average of the time from heel contact with the ground to toe leaving the ground in each gait cycle. and The pressure value of the foot is calculated by the detection element installed on the sole of the humanoid robot's foot.

[0012] In one embodiment of the present invention, the auxiliary performance indicators mainly include the following indicators and calculation methods: a. Average joint torque per unit mass of each joint during a single gait cycle: (1) Phase normalization:

[0013] in, : Normalized gait phase (0~1, corresponding to 0~100%); Phase point number, =0, ..., -1; : The number of phase points after discretizing the gait period (corresponding to 0~100%). (2) Gait mean moment curve:

[0014] in, : No. A complete gait cycle, =1, ..., K ; The total number of gait cycles used for averaging; : No. Each joint in phase Gait mean torque at the point; : No. In the 1st gait cycle, the 1st Each joint in phase Joint torque at the point; (3) Quality normalization:

[0015] in, The total mass of the robot or the object being tested; : No. Each joint in phase The average gait torque per unit mass at a given location; b. Average peak joint torque per unit mass of each joint during a single gait cycle: (1) Peak joint torque in a single gait cycle:

[0016] in, Joint number; : No. A complete gait cycle, =1, ..., K ; : Normalized gait phase, range 0≤s≤1; : No. In the 1st gait cycle, the 1st Each joint in phase Joint torque at the point; : No. In the 1st gait cycle, the 1st Peak joint torque of each joint; (2) Multi-cycle average peak joint torque:

[0017] in, : The number of gait cycles used for statistical analysis; : No. Peak joint torque of each joint after averaging over multiple gait cycles; (3) Multi-cycle average peak joint torque:

[0018] in, The total mass of the robot or the object being tested; : No. Average peak joint torque per unit mass of each joint; c. Average instantaneous joint power per unit mass within a single gait cycle: (1) Instantaneous joint power in a single cycle:

[0019] in, : No. In the 1st gait cycle, the 1st Each joint in phase angular velocity at; : Instantaneous joint power corresponding to the phase; (2) Instantaneous power after multi-cycle phase averaging:

[0020] in, : Instantaneous joint power corresponding to the phase; (3) Average instantaneous joint power per unit mass:

[0021] in, : Average instantaneous joint power curve per unit mass during a single gait period; d. Average power per unit mass within a single-step period: (1), count first Average power over one cycle:

[0022] in, : No. In the 1st gait cycle, the 1st Average power of each joint; (2) Averaging over multiple periods:

[0023] in, Average joint power after averaging multiple cycles; (3) Average power per unit mass:

[0024] in, The first step within a single-step period Average power per unit mass of each joint.

[0025] In one embodiment of the present invention, the auxiliary performance indicators include the following indicators: e. Improve efficiency:

[0026] in, The average gait power per unit mass of the humanoid robot in zero-torque mode of the exoskeleton. The average gait power per unit mass of the humanoid robot under assisted mode is the exoskeleton. f. Gait phase analysis calculation method:

[0027]

[0028] Bilateral symmetry: The percentage of standing phase for the left and right legs was calculated separately, and the differences were compared.

[0029] In one embodiment of the present invention, the auxiliary performance indicators further include the following indicators: g. Similarity measurement based on joint angle trajectories First, define the joint angle trajectory within the gait cycle, assuming the first... The angles of the joints during the gait cycle ,in, This represents the normalized gait phase. Joint angle trajectory in gait 1:

[0030] Joint angle trajectory in gait 2:

[0031] The similarity of joint angle trajectories can be measured by mean squared error (MSE):

[0032] in, and Gait 1 and gait 2 are respectively in phase The joint angle at the location; Similarity of joint angle trajectories Then, calculate the overall similarity between gait 1 and gait 2:

[0033] in, : Similarity of overall joint angles; h. Similarity measurement based on joint torque trajectory Set gait 1 and gait 2 in the first... The torque trajectory of each joint is and ,in It is the normalized phase of the gait cycle. (1) Calculation error of joint torque trajectory: The error in the joint torque trajectory can be measured by calculating the mean square error (MSE). For the first... One joint:

[0034] in: Error in the joint torque trajectory; and These are gait 1 and gait 2 in phase. Joint torque at the point; (2) Calculate the overall similarity measure: The overall gait similarity measure is obtained by averaging the similarity measures of all joints.

[0035] in, : for the similarity of overall joint torque; (3) Normalized similarity measure: To obtain a normalized similarity measure, the error values ​​can be normalized to... An interval, denoted as

[0036] The closer the normalized similarity measure is to 0, the more similar the joint torque trajectories of the two gaits are; the closer it is to 1, the greater the difference.

[0037] In one embodiment of the present invention, the auxiliary device includes through holes and connectors. Multiple through holes are provided on the first auxiliary structure and the second auxiliary structure. The connectors pass through the through holes to connect the first auxiliary structure and the second auxiliary structure. The auxiliary device includes a first auxiliary structure and a second auxiliary structure, and the first auxiliary structure and the second auxiliary structure are detachably connected. The first auxiliary structure is connected to the exoskeleton. The first auxiliary structure has a groove, and the second auxiliary structure is arc-shaped with a protrusion in the middle. The groove engages with the waist of the humanoid robot, and the protrusion is detachably connected to the waist of the humanoid robot.

[0038] Secondly, the present invention provides a gait strategy training method for a humanoid robot, used in the aforementioned robot agent evaluation system for a lower limb exoskeleton, the training method comprising: S1: Establish a simulation environment training module; use a physics engine to build a high-fidelity simulation environment in which the humanoid robot model receives action output and generates state feedback. The simulation environment includes four types of randomized terrain, initially a plane, and gradually introduces random uneven obstacles, up and down slopes and irregular road surfaces to improve the robustness and generalization ability of the strategy. S2: Set up the gait segmentation module; the gait cycle is divided into event-driven segments based on the detection of foot-ground contact force, with single-foot heel strike as the gait initiation marker; S3: Construct a human gait database module; Construct a multi-dimensional gait information database, which not only contains gait data of healthy people with different speeds and styles, but also integrates pathological gait data of patients with different levels of hemiplegia, aiming to train humanoid robot gait strategies with different styles that can adapt to specific rehabilitation needs. S4: Establish the reward function module; the reward function consists of two parts: the imitation constraint reward function and the basic reward function; the final total reward is: =

[0039] in, Total reward function : Imitate the constrained reward function, : Basic reward function; S5: Gait strategy generation module; adopts a teacher-student dual-model structure: the two models jointly form a gait generation network that integrates imitation and reinforcement. The reinforcement learning training module uses a proximal policy optimization algorithm for iterative training. Through the closed loop of "observation-action-reward-update", the student model gradually approaches the optimal gait strategy.

[0040] In one embodiment of the present invention, the method for calculating the imitation constraint reward function is as follows: =

[0041] in, = PCC ( ， ） = PCC ( ， ） = PCC ( ， ） In the formula: Weighting coefficients related to the importance of joint rehabilitation. PCC The Pearson correlation coefficient is used. For the joint torques of humanoid robots, Joint torque in humans; For the joint angles of a humanoid robot, The angle of a person's joints; For humanoid robots, the plantar pressure, Pressure on the soles of the feet.

[0042] In one embodiment of the present invention, the basic reward function includes basic motion tracking reward, motion stability constraint, joint motion optimization, foot-ground interaction optimization, system safety constraint, energy efficiency optimization, posture coordination optimization, adaptive error scaling, and altitude tracking reward. The simulated value of the basic reward function is obtained by superimposing the above functions.

[0043] The beneficial effects of this invention are: This invention provides a universal humanoid robot exoskeleton detection device, gait strategy training method, and exoskeleton-assisted walking strategy evaluation method. By setting up an auxiliary device, the exoskeleton can be mounted on a humanoid robot. The gait strategy training method adjusts the humanoid robot's gait, enabling it to mimic the walking posture of healthy individuals or people with different levels of hemiplegia. Using the humanoid robot as the testing subject, its motion accuracy and stability are high, completely eliminating individual differences and state fluctuations in human testing, ensuring the objectivity, consistency, and repeatability of the test results. Power and energy consumption can be calculated by directly measuring the torque and motion parameters at the humanoid robot's joints. The data is accurate and reliable, far superior to indirect electromyography signals or costly metabolic consumption tests.

[0044] Simultaneously, by employing exoskeleton-assisted walking strategies, data on exoskeleton dynamic load, power and energy consumption, and gait symmetry are obtained, assisting exoskeleton developers in better testing and optimizing their products. This avoids expensive and time-consuming human testing procedures, allowing for rapid iterative testing of numerous strategies after a single calibration, significantly reducing the R&D and verification costs of exoskeleton products. Furthermore, testing can be conducted under extreme or hazardous conditions without concern for human safety. The general-purpose humanoid robot can walk actively and evaluate various exoskeletons (such as single-joint and multi-joint models), with a far wider range of applications than passive dummies. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a schematic diagram of the structure of the humanoid robot exoskeleton detection device provided by the present invention.

[0047] Figure 2 This is an exploded view of the auxiliary device provided by the present invention.

[0048] Figure 3 This is a schematic diagram of the second auxiliary structure provided by the present invention.

[0049] Figure 4 This is a schematic diagram of the first auxiliary structure provided by the present invention.

[0050] Figure 5 This is a flowchart of the humanoid robot gait strategy training method provided by the present invention.

[0051] Figure 6 This is a comparison chart of the average data of 15 joints collected from a human and a humanoid robot at 28 different speeds, provided by this invention.

[0052] Figure 7 This is a comparison chart of baseline data collected by a human and a humanoid robot at three different speeds, as provided by this invention.

[0053] In the diagram: 1. Humanoid robot; 2. Exoskeleton; 3. Auxiliary device; 31. First auxiliary structure; 311. Groove; 32. Second auxiliary structure; 321. Protrusion; 33. Through hole; 34. Connector; 4. Detection element. Detailed Implementation

[0054] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Example 1 like Figures 1 to 4 As shown, this embodiment provides a detection device based on a universal humanoid robot exoskeleton, which can use a humanoid robot to replace a human to detect relevant parameters of the exoskeleton. The detection device includes a humanoid robot 1, an exoskeleton 2, an auxiliary device 3, and a detection element 4. The exoskeleton 2 to be detected is mounted on the humanoid robot 1 through the auxiliary device 3, and the detection element 4 is mounted on the sole of the humanoid robot 1's feet, which can obtain the support period time of the humanoid robot's legs.

[0056] The auxiliary device 3 includes a first auxiliary structure 31, a second auxiliary structure 32, a through hole 33, and a connector 34. Both the first auxiliary structure 31 and the second auxiliary structure 32 are provided with through holes 33. The connector 34 passes through the through holes 33 to assemble the first auxiliary structure 31 and the second auxiliary structure 32 onto the waist of the humanoid robot 1. A groove 311 is formed on one side of the first auxiliary structure 31. The groove 311 is arc-shaped, and its size and shape are adapted to the shape and size of the waist of the humanoid robot 1. The second auxiliary structure 32 is also arc-shaped. A protrusion 321 is provided in the middle of the second auxiliary structure 32. The protrusion 321 is also provided with a through hole 33. The connector 34 uses the through hole 33 on the protrusion 321 to fix the second auxiliary structure 32 to the waist of the humanoid robot 1. After the second auxiliary structure 32 is connected to the waist of the humanoid robot 1, the first auxiliary structure 31 is connected to the second auxiliary structure 32, and the groove 311 on the first auxiliary structure 31 engages with the waist of the humanoid robot 1. The exoskeleton 2 is detachably connected to the first auxiliary structure 31, allowing the auxiliary device 3 to connect to the exoskeleton 2. The groove 311 and the second auxiliary structure 32 are designed in an arc shape to provide support for the binding of the exoskeleton 2, effectively preventing the exoskeleton 2 from slipping due to gravity. Moreover, the auxiliary structure 3 is a ring structure, which will not affect the waist twisting of the humanoid robot 1 during normal walking.

[0057] Optionally, the connector 34 is a bolt, which facilitates installation and disassembly and ensures a secure lock.

[0058] Optionally, the humanoid robot 1 has built-in sensors that can collect and record the joint angles θ of each lower limb joint (hip, knee, and ankle joints) in real time throughout the walking process. baseline (t), joint angular velocity ω baseline (t) and joint torque τ baseline(t) Data that changes over time.

[0059] Optionally, the detection element 4 can be a sports shoe or pressure pad equipped with a pressure insole, worn on the foot of the humanoid robot 1, and can establish a communication connection with the controller of the exoskeleton 2 and the main control system of the humanoid robot 1.

[0060] Example 2 like Figure 5 As shown, this embodiment provides a method for training a humanoid robot's gait strategy, using the aforementioned detection device based on a general humanoid robot exoskeleton. The specific steps are as follows: S1: Establish a simulation environment training module; utilize physics engines such as Isaacsim and Mujoco to construct a high-fidelity simulation environment in which a humanoid robot model receives action outputs and generates state feedback. The environment includes four types of randomized terrain, initially a flat surface, gradually introducing random uneven obstacles, slopes, and irregular road surfaces to improve the robustness and generalization ability of the strategy.

[0061] S2: Set up the gait segmentation module; the division of gait cycles is an event-driven segmentation based on the detection of foot-ground contact force, with the heel strike of one side as the gait initiation marker, which is more in line with the biomechanical characteristics of actual human walking.

[0062] S3: Constructing a Human Gait Database Module. This module builds a multi-dimensional gait information database, including time-series data on ground reaction force, joint angles, joint torques, joint power, as well as walking linear and angular velocities. This database not only contains gait data from healthy individuals of different speeds and styles but also integrates pathological gait data from patients with different levels of hemiplegia, aiming to train robotic gait strategies with different styles that can adapt to specific rehabilitation needs.

[0063] S4: Establish the reward function module. The reward function consists of two parts: (1) Imitation constraint reward function (special part): Extract the linear velocity and angular velocity commands given by the simulation environment within a gait cycle; find the reference gait of the corresponding type that is closest to the current state in the gait database; calculate the Pearson correlation coefficients of the robot's joint angles, torques, plantar pressures and the reference gait within a single gait, and perform a weighted average; the imitation constraint reward function is calculated as follows: =

[0064] in, = PCC ( ， ） = PCC ( ， ） = PCC ( ， ） In the formula: Weighting coefficients related to the importance of joint rehabilitation. PCC The Pearson correlation coefficient is used. For the joint torques of humanoid robots, Joint torque in humans; For the joint angles of a humanoid robot, The angle of a person's joints; For humanoid robots, the plantar pressure, Pressure on the soles of the feet.

[0065] (2) Basic reward function (traditional part) The basic reward function is obtained by superimposing the functions of basic motion tracking reward, motion stability constraint, joint motion optimization, foot-ground interaction optimization, system safety constraint, energy efficiency optimization, posture coordination optimization, adaptive error scaling, and altitude tracking reward. ; Final total reward: =

[0066] in, Total reward function : Imitate the constrained reward function, : Basic reward function.

[0067] S5: Gait strategy generation module; adopts a teacher-student dual-model structure. Teacher model: a prior strategy based on human gait data to guide initial imitation; Student model: Through reinforcement learning, students can optimize autonomously and generalize strategies under the guidance of teachers. The two models together form a gait generation network that integrates imitation and reinforcement. The reinforcement learning training module uses the proximal policy optimization (PPO) algorithm for iterative training. Through the closed loop of "observation-action-reward-update", the student model gradually approaches the optimal gait policy.

[0068] By collecting gait data from healthy individuals and patients with different levels of hemiplegia to form a database, clinical medical value can be obtained. After analyzing the gait data of humans and humanoid robots, the joint data of the human lower limbs can be directly used as a target for imitation without remapping, thus avoiding the loss of original human gait information. By using the similarity of joint angles and torques throughout the gait as a reward function, the gait of the humanoid robot pays more attention to "morphological similarity" and "style transfer", making it closer to the gait of humans.

[0069] Example 3 This embodiment provides a robot proxy evaluation system for lower limb exoskeletons, which uses the general humanoid robot exoskeleton detection device provided in Embodiment 1 and the humanoid robot gait strategy training method provided in Embodiment 2, and includes the following steps: Step 1: System Setup. First, install the detection element 4 on the soles of the humanoid robot 1. Then, install the exoskeleton 2 to be evaluated on the humanoid robot 1 using the auxiliary device 3, ensuring that the joint axis of the exoskeleton 2 is roughly aligned with the corresponding lower limb joint axis of the humanoid robot 1. Establish a communication connection between the detection element 4, the exoskeleton 2 controller, and the main control system of the humanoid robot 1. Develop a data acquisition system to collect data from all three in real time at a frequency of 100Hz or higher.

[0070] Step Two: Humanoid Robot Gait Strategy Training. The humanoid robot gait strategy training method from Example Two is used. Different walking strategies are trained according to the evaluation content. If testing assistive walking strategies for normal individuals, ordinary reinforcement learning algorithms or imitation learning can be used, referencing a normal person's walking trajectory training model as a baseline strategy. If the test content is to help hemiplegic patients recover, improve gait symmetry, or help elderly people with mobility issues walk better, the maximum joint torque on one or both sides or a specific joint can be limited by modifying the reinforcement learning reward function, or imitation learning can be performed using gait data from hemiplegic patients in the database.

[0071] Step 3: Baseline Data Acquisition. Control the humanoid robot to walk along a pre-set standard test path at a predetermined speed curve. The speed can be constant or variable, and it can walk in a straight line or turn. During this process, the exoskeleton is kept in zero-torque mode, meaning it provides no active assistance or resistance. Using the robot's built-in sensors, the joint angles θ of each lower limb joint are collected and recorded in real time throughout the walking process. baseline (t), joint angular velocity ω baseline (t), joint torque τ baseline(t) and data from the built-in IMU over time.

[0072] Step 4: Assistance Strategy Testing. Under the exact same test environment, path, and set speed curve, activate the assistance strategy of the exoskeleton 2 to be tested, and control the humanoid robot 1 to walk again with the same gait parameters as in Step 3. Similarly, collect and record the joint angles θ of each lower limb joint under this exoskeleton assistance strategy in real time. baseline (t), joint angular velocity ω baseline (t), joint torque τ baseline (t) and data from the built-in IMU over time.

[0073] Step 5: Performance Index Calculation. Based on the data collected in Steps 3 and 4, calculate multi-dimensional quantitative evaluation indicators for the exoskeleton-assisted walking strategy, including assistive efficiency. Gait symmetry and auxiliary performance indicators. All calculations are performed per unit mass, i.e., by dividing the corresponding physical quantity by the total mass of the robot body and exoskeleton ( This ensures the objectivity and universality of the evaluation criteria.

[0074] The total mass of the system ( The calculation formula is as follows:

[0075] in, Mass of the humanoid robot body, The mass of the lower limb exoskeleton worn by a humanoid robot.

[0076] (1) If the exoskeleton assessment is for mobility assistance, then the core performance evaluation is for assistive efficiency. The calculation formula is as follows:

[0077] in, The energy consumed to complete a specific task without wearing an exoskeleton. The energy consumed to perform the same task while wearing an exoskeleton. and It can be calculated directly from humanoid robot data.

[0078] A positive value indicates that the exoskeleton provides assistance and saves energy; A negative value indicates that the exoskeleton increases the robot's energy consumption, i.e., generates resistance. The larger the absolute value, the more significant the effect of assistance or resistance.

[0079] (2) If the exoskeleton is used to assess the rehabilitation effect on hemiplegic patients, the core performance evaluation is gait symmetry. The gait symmetry The calculation formula is as follows:

[0080] in, For gait symmetry, The average support period time of the left leg under multiple gait states is the average of the entire time from the heel touching the ground to the toes leaving the ground in each gait cycle. The support phase time of the right leg is the average of the time from heel contact with the ground to toe leaving the ground in each gait cycle. and The pressure value of the foot is calculated by the detection element 4 installed on the foot of the humanoid robot 1.

[0081] (3) Calculation of auxiliary performance indicators To comprehensively evaluate strategy performance, the following supplementary metrics are provided: a. Average joint torque per unit mass of each joint during a single gait cycle: (1) Phase normalization:

[0082] in, : Normalized gait phase (0~1, corresponding to 0~100%); Phase point number, =0, ..., -1; : The number of phase points after discretizing the gait period (corresponding to 0~100%).

[0083] (2) Gait mean moment curve:

[0084] in, : No. A complete gait cycle, =1, ..., K ; The total number of gait cycles used for averaging; : No. Each joint in phase Gait mean torque at the point; : No. In the 1st gait cycle, the 1st Each joint in phase Joint torque at the point.

[0085] (3) Quality normalization:

[0086] in, The total mass of the robot or the object being tested; : No. Each joint in phase The average torque per unit mass during gait at that location.

[0087] b. Average peak joint torque per unit mass of each joint during a single gait cycle: (1) Peak joint torque in a single gait cycle:

[0088] in, Joint number; : No. A complete gait cycle, =1, ..., K ; : Normalized gait phase, range 0≤s≤1; : No. In the 1st gait cycle, the 1st Each joint in phase Joint torque at the point; : No. In the 1st gait cycle, the 1st Peak joint torque of each joint.

[0089] (2) Multi-cycle average peak joint torque:

[0090] in, : The number of gait cycles used for statistical analysis; : No. The peak joint torque of each joint after averaging over multiple gait cycles.

[0091] (3) Multi-cycle average peak joint torque:

[0092] in, The total mass of the robot or the object being tested; : No. Average peak joint torque per unit mass of each joint.

[0093] c. Average instantaneous joint power per unit mass within a single gait cycle: (1) Instantaneous joint power in a single cycle:

[0094] in, : No. In the 1st gait cycle, the 1st Each joint in phase angular velocity at; : Instantaneous joint power corresponding to the phase.

[0095] (2) Instantaneous power after multi-cycle phase averaging:

[0096] in, : Instantaneous joint power corresponding to the phase.

[0097] (3) Average instantaneous joint power per unit mass:

[0098] in, : Average instantaneous joint power curve per unit mass during single gait period.

[0099] d. Average power per unit mass within a single-step period: (1), count first Average power over one cycle:

[0100] in, : No. In the 1st gait cycle, the 1st Average power of each joint.

[0101] (2) Averaging over multiple periods:

[0102] in, : Average joint power after averaging over multiple cycles.

[0103] (3) Average power per unit mass:

[0104] in, The first step within a single-step period Average power per unit mass of each joint.

[0105] e. Improve efficiency:

[0106] in, The average gait power per unit mass of the humanoid robot in zero-torque mode of the exoskeleton. This represents the average gait power per unit mass of the humanoid robot under assisted mode.

[0107] f. Gait phase analysis calculation method:

[0108]

[0109] Bilateral symmetry: The percentage of standing phase for the left and right legs was calculated separately, and the differences were compared.

[0110] g. Similarity measurement based on joint angle trajectories First, define the joint angle trajectory within the gait cycle, assuming the first... The angles of the joints during the gait cycle ,in, This represents the normalized gait phase. Joint angle trajectory in gait 1:

[0111] Joint angle trajectory in gait 2:

[0112] The similarity of joint angle trajectories can be measured by mean squared error (MSE):

[0113] in, and Gait 1 and gait 2 are respectively in phase The joint angle at the location; Similarity of joint angle trajectories Then, calculate the overall similarity between gait 1 and gait 2:

[0114] in, : refers to the similarity of the overall joint angles.

[0115] h. Similarity measurement based on joint torque trajectory Set gait 1 and gait 2 in the first... The torque trajectory of each joint is and ,in It is the normalized phase of the gait cycle. (1) Calculation error of joint torque trajectory: The error in the joint torque trajectory can be measured by calculating the mean square error (MSE). For the first... One joint:

[0116] in: Error in the joint torque trajectory; and These are gait 1 and gait 2 in phase. Joint torque at the point.

[0117] (2) Calculate the overall similarity measure: The overall gait similarity measure is obtained by averaging the similarity measures of all joints.

[0118] in, : This refers to the similarity of the overall joint torque.

[0119] (3) Normalized similarity measure: To obtain a normalized similarity measure, the error values ​​can be normalized to... An interval, denoted as

[0120] The closer the normalized similarity measure is to 0, the more similar the joint torque trajectories of the two gaits are; the closer it is to 1, the greater the difference.

[0121] Example 4 like Figures 6 to 7 As shown, this embodiment provides a hip joint assistive exoskeleton, and uses an exoskeleton assistive strategy evaluation method based on a general humanoid robot provided in Embodiment 3 to evaluate the hip joint assistive exoskeleton.

[0122] Step 1: System Setup. Select a general-purpose humanoid robot (such as the Unitree G1 or a similar model). Equip the humanoid robot with athletic shoes featuring insoles with pressure sensing elements. Securely attach the hip exoskeleton (or other types of exoskeletons, but we'll use the hip exoskeleton as an example) to the torso and legs of the humanoid robot 1 using auxiliary device 3. Align the hip joint hinge of the exoskeleton 2 with the robot's hip joint axis. Connect the pressure insoles, exoskeleton controller, and the humanoid robot's main control computer. Develop a data acquisition system to collect data from all three in real time at a frequency of 100Hz or higher.

[0123] Step 2: Train the general humanoid robot from Step 1 using the humanoid robot gait strategy training method in Example 2.

[0124] Step 3: Baseline Data Acquisition. The humanoid robot is programmed to walk in a straight line for 20 seconds at different speeds on a flat indoor surface. Joint data is recorded at 100Hz via the humanoid robot's API interface. The gait is divided into left and right sides based on the right heel strike time. The divided gait is normalized to 0-100%, and acceleration / deceleration components are removed. The intermediate gait data are superimposed and averaged to calculate the gait curve of the humanoid robot under this control strategy for comparison. Baseline data of the lower limbs and lumbar joints collected from the human and humanoid robot at 28 different speeds are shown below. Figure 6 As shown.

[0125] Step 4: Assist Strategy Testing: Keeping all conditions unchanged, equip the humanoid robot with an exoskeleton and activate the exoskeleton's assist controller. Control the robot to walk again and record joint data (joint angle θ) under different assist modes. assist (t), joint angular velocity ω assist (t) and joint torque τ assist (t)).

[0126] like Figure 7 The image shows a comparison of the angles, torques, and power of the right lower limb hip, knee, and ankle joints at three speeds (0.5 m / s, 1.0 m / s, and 1.5 m / s) for a human and a humanoid robot using a gait strategy based on imitation and reinforcement learning at the same three speeds (0.5 m / s, 1.0 m / s, and 1.5 m / s). Figure 7 As can be seen, the gait of humanoid robots is very similar to that of humans, especially in terms of joint angles. Both the angles themselves and their trends with speed show a high degree of consistency. Furthermore, these similarities can be further improved by modifying the model training parameters. Although there are some differences in joint torque and power during human-humanoid robot walking, the main reason lies in the different actuation methods. For the evaluation of assistive exoskeletons, the focus is more on reducing joint torque and power, rather than their absolute values.

[0127] Step 5: Performance Calculation: After completing baseline and assistive strategy testing, the recorded data is processed and analyzed to comprehensively and quantitatively evaluate the performance of the exoskeleton assistive walking strategy. Some energy-related calculations are performed per unit mass to ensure the objectivity of the evaluation criteria. The calculations mainly focus on multiple dimensions such as dynamic load, power and energy consumption, and gait symmetry. This helps exoskeleton developers better test and understand their exoskeleton assistive walking or rehabilitation strategies.

[0128] In summary, this invention provides a universal humanoid robot exoskeleton detection device, gait strategy training method, and exoskeleton-assisted walking strategy evaluation method. By setting up an auxiliary device 3, the exoskeleton can be mounted on the humanoid robot. The gait strategy training method adjusts the humanoid robot's gait, enabling it to mimic the walking posture of healthy individuals or people with different levels of hemiplegia. The humanoid robot, as the testing subject, exhibits high motion accuracy and stable state, completely eliminating individual differences and state fluctuations in human testing, ensuring the objectivity, consistency, and repeatability of the test results. Power and energy consumption can be calculated by directly measuring the torque and motion parameters at the humanoid robot's joints, providing accurate and reliable data that far surpasses indirect electromyography signals or costly metabolic consumption tests.

[0129] Simultaneously, by employing exoskeleton-assisted walking strategies, data on exoskeleton dynamic load, power and energy consumption, and gait symmetry are obtained, assisting exoskeleton developers in better testing and optimizing their products. This avoids expensive and time-consuming human testing procedures, allowing for rapid iterative testing of numerous strategies after a single calibration, significantly reducing the R&D and verification costs of exoskeleton products. Furthermore, testing can be conducted under extreme or hazardous conditions without concern for human safety. The general-purpose humanoid robot can walk actively and evaluate various exoskeletons (such as single-joint and multi-joint models), with a far wider range of applications than passive dummies.

[0130] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A robotic proxy evaluation system for lower limb exoskeletons, characterized in that, The invention includes a humanoid robot exoskeleton detection device, the detection device comprising: a humanoid robot (1), an exoskeleton (2), an auxiliary device (3) and a detection element (4), wherein the exoskeleton (2) is connected to the humanoid robot (1) through the auxiliary device (3), and the detection element (4) is installed on the sole of the foot of the humanoid robot (1); The robotic proxy evaluation system for the lower limb exoskeleton includes the following steps: Step 1: System setup; First, install the detection element (4) on the sole of the foot of the humanoid robot (1), install the exoskeleton (2) to be evaluated on the humanoid robot (1) through the auxiliary device (3), and ensure that the joint axis of the exoskeleton (2) is roughly aligned with the corresponding lower limb joint axis of the humanoid robot (1), and establish a communication connection between the detection element (4), the exoskeleton (2) controller and the main control system of the humanoid robot (1), and develop a data acquisition system to collect the data of the three in real time; Step 2: Gait strategy training; train different walking strategies for the humanoid robot according to the assessment content; if the test is to test the assistive walking strategy of normal people, the walking trajectory training model of normal people is used as the benchmark strategy; if the test content is to help hemiplegic patients recover, improve gait symmetry or assist walking, the maximum joint torque of unilateral and bilateral joints is limited by modifying the reinforcement learning reward function, and the gait data of hemiplegic patients in the database is used for imitation learning. Step 3: Baseline data acquisition; control the humanoid robot (1) to walk on a preset standard test path at a set speed curve; during this process, keep the exoskeleton (2) in zero torque mode; through the built-in sensors of the humanoid robot (1), collect and record the joint angle θ of each lower limb joint in real time during the entire walking process. baseline Joint angular velocity ω baseline Joint torque τ baseline And data from the built-in IMU that changes over time; Step 4: Assistance strategy test; Under the same test environment, path and set speed curve, activate the assist strategy to be tested of the exoskeleton (2), and control the humanoid robot (1) to walk again with the same gait parameters as in Step 3. Similarly, collect and record the joint angles θ of each lower limb joint under the exoskeleton assist strategy in real time. baseline Joint angular velocity ω baseline Joint torque τ baseline And data from the built-in IMU that changes over time; Step 5: Performance Index Calculation; Based on the data collected in Steps 3 and 4, calculate multi-dimensional quantitative evaluation indicators for the exoskeleton-assisted walking strategy, including assistive efficiency. Gait symmetry and auxiliary performance indicators; some energy consumption-related performance indicators are calculated per unit mass, that is, by dividing the corresponding physical quantity by the total mass of the humanoid robot body and exoskeleton. This is to ensure the objectivity and universality of the evaluation criteria.

2. The robot proxy evaluation system for lower limb exoskeletons according to claim 1, characterized in that, If used to evaluate the mobility assistance capabilities of the exoskeleton (2), the evaluation metric is the assist efficiency. The aforementioned assist efficiency The calculation formula is as follows: in, The energy consumed to complete a specific task without wearing an exoskeleton. The energy consumed to perform the same task while wearing an exoskeleton. and It can be calculated directly from humanoid robot data.

3. The robot proxy evaluation system for lower limb exoskeletons according to claim 1, characterized in that, If the exoskeleton (2) is used to evaluate the rehabilitation effect on hemiplegic patients, then the core performance evaluation is gait symmetry. The gait symmetry The calculation formula is as follows: in, For gait symmetry, The average support period time of the left leg under multiple gait states is the average time from the heel touching the ground to the toes leaving the ground in each gait cycle. The support phase time of the right leg is the average of the time from heel contact with the ground to toe leaving the ground in each gait cycle. and The pressure value of the foot is calculated by the detection element (4) installed on the sole of the foot of the humanoid robot (1).

4. The robot proxy evaluation system for lower limb exoskeletons according to claim 1, characterized in that, The auxiliary performance indicators include the following indicators and calculation methods: a. Average joint torque per unit mass of each joint during a single gait cycle: (1) Phase normalization: in, Normalized gait phase; Phase point number, =0, ..., -1; The number of phase points after discretizing the gait period; (2) Gait mean moment curve: in, : No. A complete gait cycle, =1, ..., K ; The total number of gait cycles used for averaging; : No. Each joint in phase Gait mean torque at the location; : No. In the 1st gait cycle, the 1st Each joint in phase Joint torque at the point; (3) Quality normalization: in, The total mass of the robot or the object being tested; : No. Each joint in phase The average gait torque per unit mass at the location; b. Average peak joint torque per unit mass of each joint during a single gait cycle: (1) Peak joint torque in a single gait cycle: in, Joint number; : No. A complete gait cycle, =1, ..., K ; : Normalized gait phase, range 0≤s≤1; : No. In the 1st gait cycle, the 1st Each joint in phase Joint torque at the point; : No. In the 1st gait cycle, the 1st Peak joint torque of each joint; (2) Multi-cycle average peak joint torque: in, : The number of gait cycles used for statistical analysis; : No. Peak joint torque of each joint after averaging over multiple gait cycles; (3) Multi-cycle average peak joint torque: in, The total mass of the robot or the object being tested; : No. Average peak joint torque per unit mass of each joint; c. Average instantaneous joint power per unit mass within a single gait cycle: (1) Instantaneous joint power in a single cycle: in, : No. In the 1st gait cycle, the 1st Each joint in phase angular velocity at; : Instantaneous joint power corresponding to the phase; (2) Instantaneous power after multi-cycle phase averaging: in, : Instantaneous joint power corresponding to the phase; (3) Average instantaneous joint power per unit mass: in, : Average instantaneous joint power curve per unit mass during a single gait period; d. Average power per unit mass within a single-step period: (1), count first Average power over one cycle: in, : No. In the 1st gait cycle, the 1st Average power of each joint; (2) Averaging over multiple periods: in, Average joint power after averaging multiple cycles; (3) Average power per unit mass: in, The first step within a single-step period Average power per unit mass of each joint.

5. The robot proxy evaluation system for lower limb exoskeletons according to claim 4, characterized in that, The auxiliary performance indicators include the following: e. Improve efficiency: in, The average gait power per unit mass of the humanoid robot in zero-torque mode of the exoskeleton. The average gait power per unit mass of the humanoid robot under assisted mode is the exoskeleton. f. Gait phase analysis calculation method: Bilateral symmetry: The percentage of standing phase for the left and right legs was calculated separately, and the differences were compared.

6. The robot proxy evaluation system for lower limb exoskeletons according to claim 5, characterized in that, The auxiliary performance indicators also include the following: g. Similarity measurement based on joint angle trajectories First, define the joint angle trajectory within the gait cycle, assuming the first... The angles of the joints during the gait cycle ,in, This represents the normalized gait phase. Joint angle trajectory in gait 1: Joint angle trajectory in gait 2: The similarity of joint angle trajectories can be measured by mean squared error (MSE): in, and Gait 1 and gait 2 are respectively in phase The joint angle at the location; Similarity of joint angle trajectories Then, calculate the overall similarity between gait 1 and gait 2: in, : Similarity of overall joint angles; h. Similarity measurement based on joint torque trajectory Set gait 1 and gait 2 in the first... The torque trajectory of each joint is and ,in It is the normalized phase of the gait cycle. (1) Calculation error of joint torque trajectory: The error in the joint torque trajectory can be measured by calculating the mean square error (MSE). For the first... One joint: in: Error in the joint torque trajectory; and These are gait 1 and gait 2 in phase. Joint torque at the point; (2) Calculate the overall similarity measure: The overall gait similarity measure is obtained by averaging the similarity measures of all joints. in, : for the similarity of overall joint torque; (3) Normalized similarity measure: To obtain a normalized similarity measure, the error values ​​can be normalized to... An interval, denoted as The closer the normalized similarity measure is to 0, the more similar the joint torque trajectories of the two gaits are; the closer it is to 1, the greater the difference.

7. The robot proxy evaluation system for lower limb exoskeletons according to claim 1, characterized in that, The auxiliary device (3) includes through holes (33) and connectors (34). Multiple through holes (33) are provided on the first auxiliary structure (31) and the second auxiliary structure (32). The connectors (34) pass through the through holes (33) to connect the first auxiliary structure (31) and the second auxiliary structure (32). The auxiliary device (3) includes a first auxiliary structure (31) and a second auxiliary structure (32). The first auxiliary structure (31) and the second auxiliary structure (32) are detachably connected. The first auxiliary structure (31) is connected to the exoskeleton (2). The first auxiliary structure (31) has a groove (311), and the second auxiliary structure (32) is arc-shaped with a protrusion (321) in the middle. The groove (311) is engaged with the waist of the humanoid robot (1), and the protrusion (321) is detachably connected to the waist of the humanoid robot (1).

8. A method for training gait strategies for a humanoid robot, characterized in that, The training method for a robotic agent evaluation system for a lower limb exoskeleton as described in any one of claims 1-7 includes: S1: Establish a simulation environment training module; use a physics engine to build a high-fidelity simulation environment in which the humanoid robot (1) model receives action output and generates state feedback. The simulation environment includes four types of randomized terrain, initially a plane, and gradually introduces random bumps, slopes and irregular roads to improve the robustness and generalization ability of the strategy. S2: Set up the gait segmentation module; the gait cycle is divided into event-driven segments based on the detection of foot-ground contact force, with single-foot heel strike as the gait initiation marker; S3: Construct a human gait database module; Construct a multi-dimensional gait information database, which not only contains gait data of healthy people with different speeds and styles, but also integrates pathological gait data of patients with different levels of hemiplegia, aiming to train humanoid robot gait strategies with different styles that can adapt to specific rehabilitation needs. S4: Establish the reward function module; the reward function consists of two parts: the imitation constraint reward function and the basic reward function; the final total reward is: = in, Total reward function, : Imitate the constrained reward function, : Basic reward function; S5: Gait strategy generation module; adopts a teacher-student dual-model structure: the two models jointly form a gait generation network that integrates imitation and reinforcement. The reinforcement learning training module uses a proximal policy optimization algorithm for iterative training. Through the "observation-action-reward-update" closed loop, the student model gradually approaches the optimal gait strategy.

9. The gait strategy training method for a humanoid robot according to claim 8, characterized in that, The method for calculating the imitation constraint reward function is as follows: = in, = PCC ( ， ） = PCC ( ， ） = PCC ( ， ） In the formula: Weighting coefficients related to the importance of joint rehabilitation. PCC The Pearson correlation coefficient is used. For the joint torques of humanoid robots, Joint torque in humans; For the joint angles of a humanoid robot, The angle of a person's joints; For humanoid robots, the plantar pressure, Pressure on the soles of the feet.

10. A gait strategy training method for a humanoid robot according to claim 9, characterized in that, The basic reward function The basic reward function is obtained by superimposing the above functions, which include basic motion tracking reward, motion stability constraint, joint motion optimization, foot-ground interaction optimization, system safety constraint, energy efficiency optimization, posture coordination optimization, adaptive error scaling, and altitude tracking reward.

Images