Multimodal environment perception and adaptive chassis control method for humanoid robots

Abstract

The present invention relates to the field of robot intelligent control technology, specifically, the present invention relates to a multimodal environmental perception and adaptive chassis control method for a humanoid robot. The present invention monitors the contact force vector and sliding trend in real time through a multidimensional tactile sensor, maps the posture drift to the chassis coordinate system based on a spatiotemporal synchronization layer constrained by Lie group, generates a compensation vector, adopts a three-level tactile response layer to dynamically switch the force / position mixed impedance mode, combines the omnidirectional chassis reverse translation to offset the drift, and optimizes the compensation parameters through a deep deterministic policy gradient algorithm. The system ensures the timing alignment of control instructions through a spatiotemporal stamp synchronization engine, uses an inertial measurement unit to suppress posture oscillation during the compensation process, constructs an obstacle avoidance interference domain based on the kinematic chain of the operating arm, and uses a B-spline curve to smooth the compensation trajectory, thereby improving the accuracy, stability and safety of the humanoid robot in complex contact operations, and providing an efficient and reliable solution for human-machine collaboration scenarios.

Description

Technical Field

[0001] The present invention relates to the field of robot intelligent control technology, and in particular to a multimodal environment perception and adaptive chassis control method for a humanoid robot. Background Art

[0002] Robot intelligent control is an important technology. In complex operating scenarios, humanoid robots achieve precise operation, dynamic obstacle avoidance and stable walking through precise coordination of environmental perception and chassis movement. As the demand for human-machine collaboration increases, robots face the challenges of force disturbance and posture drift in contact operations.

[0003] However, traditional control technologies suffer from core issues of multimodal perception lag and chassis-robot coordination failure. Existing solutions rely on single vision or force sensors and lack a mechanism for the spatiotemporal synchronization of tactile, force, and posture data. This results in delayed responses to contact force disturbances exceeding the set time. In precision assembly, when the end-of-operation contact force exceeds the safety threshold, the posture drift is not promptly mapped to the chassis coordinate system. The robot slips due to single-axis overload, significantly increasing positioning errors. Traditional force / position control mode switching logic is rigid, and position-dominated control is still used during the critical sliding phase, resulting in sudden contact force changes that damage components and, in turn, reduce part yield. Furthermore, the chassis motion is not coordinated with the robot's kinematic chain, and the compensation path is prone to falling into the robot's restricted working range, causing joint interference and increasing system failure rate. This chain reaction of perception lag and coordination failure ultimately reduces the success rate of traditional robots in complex contact operations, making them difficult to adapt to high-precision human-robot collaboration scenarios. To address this technical issue, we propose a multimodal environmental perception and adaptive chassis control method for humanoid robots. Summary of the Invention

[0004] The purpose of the present invention is to provide a multimodal environment perception and adaptive chassis control method for a humanoid robot to solve the problems raised in the above background technology.

[0005] 1. Due to the lag in multimodal perception in traditional solutions, the response delay to contact force disturbances exceeds the set time, resulting in large positioning errors. Therefore, this case uses real-time monitoring with multi-dimensional tactile sensors, combined with spatiotemporal synchronization layer mapping of posture drift to generate chassis compensation vectors, which can shorten response delays and accurately control positioning errors.

[0006] 2. Due to the failure of traditional chassis and robotic arm coordination, the compensation path is prone to falling into restricted areas and causing interference. Therefore, this case constructs an obstacle avoidance interference domain based on the kinematic chain and uses B-spline curves to smooth the compensation trajectory, which can automatically avoid the working restricted area and reduce the system failure rate.

[0007] To achieve the above objectives, one of the objectives of the present invention is to provide a method for multimodal environment perception and adaptive chassis control of a humanoid robot, comprising the following steps:

[0008] S1. Use a multi-dimensional tactile sensor to monitor the contact force vector and sliding trend of the end of the operating mechanism in real time. When the contact force vector exceeds a preset safety threshold or the axial deviation rate of the sliding trend is greater than 5 mm per second, activate the collaborative compensation mechanism;

[0009] S2. The collaborative compensation mechanism maps the posture drift of the end of the operating mechanism due to the contact force disturbance to the chassis motion coordinate system based on the spatiotemporal synchronization layer of Lie group constraints, and generates the direction and amplitude of the chassis compensation vector;

[0010] S3. Dual-system collaborative compensation is performed based on the chassis compensation vector. Dual-system collaborative compensation includes operating mechanism control and chassis motion control. The operating mechanism control dynamically switches the force / position hybrid impedance mode according to the sliding trend. That is, position-dominant impedance control is adopted in the stable contact phase, and force-tracking impedance control is switched in the critical sliding phase. The chassis motion control omnidirectionally moves the chassis in the direction opposite to the operating mechanism movement direction, offsetting the posture drift in real time and maintaining the positioning error of the end in the working coordinate system.

[0011] S4. Use the deep deterministic policy gradient algorithm to optimize the mapping relationship between the sliding trend judgment threshold and the chassis compensation vector amplitude coefficient.

[0012] As a further improvement of the present technical solution, the sliding trend dynamic switching force / position mixed impedance mode includes:

[0013] Construct a three-level tactile response layer, where the primary response layer analyzes the contact force vector direction, the intermediate response layer calculates the sliding trend probability distribution, and the advanced response layer combines the two to generate impedance switching instructions;

[0014] When the sliding tendency probability is greater than the preset threshold, the control switches from position-dominant impedance control to force-tracking impedance control, and the chassis compensation collaborative flag is activated.

[0015] As a further improvement of this technical solution, between the operating mechanism control and the chassis motion control:

[0016] The tactile signal and chassis posture data are timestamped through the time-space stamp synchronization engine, and dynamic compensation is performed based on the timestamp deviation. If the deviation between the operating mechanism control instruction and the chassis execution is greater than 2 milliseconds, an empty instruction cycle is inserted to align the timing.

[0017] As a further improvement of this technical solution, the chassis monitors the compensation reaction force in real time when it moves in the opposite direction, and the attitude oscillation amplitude during the compensation process is fed back through the chassis inertial measurement unit. If the attitude oscillation amplitude is greater than Then a reverse compensation vector correction component is generated to offset the secondary disturbance.

[0018] As a further improvement of this technical solution, the determination of the critical sliding stage includes:

[0019] A sliding energy integration model is established to integrate the tangential acceleration of the contact surface of the operating end. When the integral value exceeds the threshold associated with the material friction coefficient, the critical sliding response is triggered 50 milliseconds in advance. When triggered, the contact force of the operating end is reduced to suppress the accumulation of sliding kinetic energy.

[0020] As a further improvement of the present technical solution, an obstacle avoidance interference domain is constructed based on the kinematic chain of the manipulator arm during the chassis translation motion, and the restricted area of ​​the manipulator workspace is automatically avoided in the compensation path. A B-spline curve is used to smooth the compensation trajectory to control the chassis acceleration continuously and without sudden changes.

[0021] As a further improvement of this technical solution, a collaborative arbitrator is set up. When the impedance control output of the operating mechanism conflicts with the chassis compensation vector, the control weight is redistributed with the goal of minimizing the end positioning error, and a conflict arbitration strategy is formulated to prioritize the execution of the chassis compensation vector and switch the operating mechanism to passive compliance mode.

[0022] As a further improvement of the present technical solution, the operation when the force tracking impedance control is executed is as follows:

[0023] The stiffness characteristics of the contact object are identified through high-frequency spectrum analysis of the tactile signal, and the force tracking gain is dynamically adjusted. Objects that withstand a force of more than 100 Newtons per millimeter of length are recorded as high-stiffness objects, and low-gain slow tracking is used for high-stiffness objects. Objects that withstand a force of less than 30 Newtons per millimeter of length are recorded as low-stiffness objects, and high-gain fast tracking is used for low-stiffness objects.

[0024] As a further improvement of this technical solution, fourteen compensation trajectories are rendered in the virtual coordinate system of the end of the operating mechanism to form a posture error heat map. When the posture error heat map shows local error accumulation, the compensation vector weight of the area is automatically enhanced.

[0025] As a further improvement of this technical solution, the deep deterministic policy gradient algorithm optimization process includes:

[0026] A compensation vector-task scenario association knowledge base is constructed to store the historical optimal compensation parameters of different operation tasks. When the similarity between the detected task scenario and the historical task scenarios in the compensation vector-task scenario association knowledge base reaches a similarity threshold, the compensation vector is loaded with the parameters from the compensation vector-task scenario association knowledge base to initialize the compensation vector.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] The present invention captures contact force vectors and sliding trends in real time through the fusion mechanism of multi-dimensional tactile sensors and spatiotemporal synchronization layers, and quickly maps posture drift to the chassis coordinate system based on Lie group constraints, thereby reducing the contact force disturbance response delay time and positioning error, and solving the problem of insufficient operation accuracy caused by the lag of traditional single-modal perception. Through the three-level tactile response layer and dual-system collaborative compensation technology, position-dominant impedance control is adopted in the stable contact stage, and force tracking impedance control is switched in the critical sliding stage. The chassis reverse translation is combined to offset the drift, and the deep deterministic strategy gradient algorithm is used to optimize the compensation parameter mapping, thereby improving the part qualification rate and reducing the risk of component damage caused by sudden changes in contact force. By constructing an obstacle avoidance interference domain and B-spline curve trajectory smoothing technology, the robot arm's working restricted area is automatically avoided during the chassis compensation movement, reducing the system failure rate. At the same time, the inertial measurement unit feedback is used to correct the posture oscillation to ensure that the chassis acceleration is continuous and without sudden changes during the compensation process, thereby improving the movement safety and trajectory smoothness in complex contact operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is the overall workflow diagram of the present invention. DETAILED DESCRIPTION

[0030] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0031] See also Figure 1 As shown, this embodiment provides a method for multimodal environment perception and adaptive chassis control of a humanoid robot, comprising the following steps:

[0032] When a humanoid robot performs contact operations, in order to perceive the force state of the end-user in real time and avoid overload risks, a dynamic monitoring and compensation triggering mechanism must be established using multi-dimensional tactile sensors. The specific implementation is as follows:

[0033] S1. A multi-dimensional tactile sensor and a fiber Bragg grating sliding tendency sensor are integrated at the end of the operating mechanism. The former collects the X, Y, and Z-axis force and torque components in real time at a frequency of 1000 Hz, and the latter calculates the axial offset rate by measuring the grating wavelength offset. For example, when the end grasps the workpiece, the six-dimensional force sensor continuously monitors the vertical force on the Z axis. If the force value exceeds the preset safety threshold within five consecutive sampling cycles, it is determined that there is a risk of contact force overload. At the same time, if the fiber Bragg grating sensor detects that the axial offset rate is greater than 5 mm / s for three consecutive cycles, it is determined that the sliding tendency exceeds the standard. The contact force vector and sliding tendency of the end of the operating mechanism are monitored in real time by the multi-dimensional tactile sensor. When the contact force vector exceeds the preset safety threshold or the axial offset rate of the sliding tendency is greater than 5 mm / s, the collaborative compensation mechanism is activated.

[0034] When any component of the contact force vector exceeds the safety threshold or the sliding trend rate exceeds the standard, the system triggers the collaborative compensation mechanism through the hard real-time controller. To avoid false triggering of a single sensor, the "OR gate" logic is set as follows:

[0035] If any of the conditions of force overload or sliding exceeds the limit is met, compensation is activated immediately. For example, if the end sliding rate reaches 6mm / s (exceeding the threshold of 5mm / s) due to the smooth surface of the workpiece, even if the contact force does not reach the overload threshold, compensation will still be triggered to ensure a rapid response to potential sliding risks. After activating compensation, the position control command of the current operation end is first frozen to prevent overload from being aggravated by continuous force application, and mechanical vibration interference is eliminated through the tactile signal preprocessing module based on the Kalman filter algorithm to ensure the accuracy of the triggering conditions. For example, when the force sensor has an instantaneous spike signal due to vibration, the filtering algorithm eliminates outliers by fitting historical data, effectively avoiding the risk of workpiece damage or robot loss of control caused by the hysteresis detection of traditional single sensors.

[0036] After activating the collaborative compensation mechanism, in order to achieve precise coordination between the end-operator and chassis motion, the end-operator pose drift needs to be mapped to the chassis coordinate system through the spatiotemporal synchronization layer to generate a precise compensation vector. The specific implementation is as follows:

[0037] S2. When the posture of the operating end changes due to contact force disturbance, the three-dimensional translation error and three-dimensional rotation error are first solved in real time by the inertial measurement unit installed at the end. In order to eliminate the error caused by the coordinate system difference, based on the Lie group constraint theory, a special Euclidean group is used to describe the posture relationship between the end coordinate system and the chassis coordinate system. The six-dimensional posture error is converted into a vector representation in the Lie algebra space through the Lie group index mapping. Then, with the help of the pre-calibrated hand-eye relationship matrix, the end error is mapped from the local coordinate system to the chassis global motion coordinate system to ensure the geometric consistency of the time and space transformation. For example, the end rotates around the Z axis. The drift can be converted into an equivalent rotation in the chassis coordinate system after Lie group operation, avoiding compensation failure caused by coordinate conversion deviation. According to the mapped posture error, the direction and amplitude of the compensation vector are calculated in combination with the chassis kinematic model. For translation error, the three-dimensional translation error is directly used as the compensation displacement of the chassis in the X, Y, and Z axes. For rotation error, the speed difference of each wheel of the chassis is solved by inverse kinematics to offset the angle deviation. In order to improve the smoothness of compensation, the B-spline curve is used to fit the compensation trajectory to ensure that the chassis acceleration is continuous and without mutation. At the same time, an obstacle avoidance interference domain is constructed based on the kinematic chain of the manipulator arm to automatically avoid the working restricted area of ​​the robotic arm. For example, when the compensation path involves the extreme position of the robotic arm joint, the system adjusts the direction of the compensation vector in real time to make the chassis move along a safe path to avoid the risk of joint interference.

[0038] The LiDAR installed on the chassis monitors the compensated posture error in real time. If the residual error exceeds the allowable range, secondary compensation is triggered. Specifically:

[0039] The extended Kalman filter algorithm is used to fuse the tactile signal and chassis posture data to form a closed-loop feedback control, and the final posture error is controlled within For example, when there is still an X-axis translation error of 0.3mm after the first compensation, the system automatically generates an additional compensation vector, that is, , the driving chassis completes secondary adjustment, which improves the accuracy compared with traditional open-loop compensation;

[0040] Through the above-mentioned spatiotemporal synchronization mechanism based on Lie group constraints, the system realizes the full-process mathematical modeling and dynamic calibration from end disturbance to chassis compensation, reduces the motion coupling error between the operating mechanism and the chassis, effectively solves the compensation delay problem caused by coordinate transformation lag in traditional collaborative control, and enhances the motion accuracy and safety of humanoid robots in complex contact scenarios.

[0041] S3. Dual-system collaborative compensation is performed based on the chassis compensation vector. Dual-system collaborative compensation includes operating mechanism control and chassis motion control. The operating mechanism control dynamically switches the force / position hybrid impedance mode according to the sliding trend. That is, position-dominant impedance control is adopted in the stable contact phase, and force-tracking impedance control is switched in the critical sliding phase. The chassis motion control omnidirectionally moves the chassis in the direction opposite to the operating mechanism movement direction, offsetting the posture drift in real time and maintaining the positioning error of the end in the working coordinate system.

[0042] After completing real-time monitoring of the contact force and sliding trend of the operating mechanism end, in order to achieve refined control of the contact state, the system constructs a three-level tactile response layer to achieve dynamic switching of the force / position hybrid impedance mode. The specific implementation is as follows:

[0043] Sliding trend dynamic switching force / position mixed impedance mode includes:

[0044] Construct a three-level tactile response layer, where the primary response layer analyzes the contact force vector direction, the intermediate response layer calculates the sliding trend probability distribution, and the advanced response layer combines the two to generate impedance switching instructions;

[0045] In order to analyze the tactile signals layer by layer and generate precise control instructions, the primary response layer uses a vector decomposition algorithm to decompose the six-dimensional contact force signal into normal force and tangential force components, identify the direction of the contact force vector, and determine that the normal force accounts for more than 70% if it is a stable contact, and the tangential force accounts for more than 50% if it is a sliding trend. The analysis frequency reaches 100Hz to ensure real-time performance. The intermediate response layer models the tangential force fluctuation sequence based on the hidden Markov model and calculates the probability distribution of sliding within the next 50ms. The probability value range is For example, when the tangential force exceeds 30% of the normal force for three consecutive cycles, the slip probability increases to over 60%. The advanced response layer fuses the force vector direction and the slip probability through the fuzzy logic controller to generate an impedance mode switching instruction. For example, when the slip probability is greater than 0.7 and the tangential force direction is along the positive direction of the x-axis, the "switch to force tracking control" signal is output. The system presets the sliding trend probability threshold to 0.6. When the intermediate response layer output probability is greater than 0.6, the mode switch is triggered:

[0046] Switch from position-dominant impedance control to force-tracking impedance control, and send a collaborative flag to the chassis control system at the same time to activate the chassis compensation program. In position-dominant mode, when the end position error is less than 0.2mm, maintain stiffness control to ensure positioning accuracy. In force tracking mode, the contact force tracking error is made less than 5N through the proportional-integral controller. For example, when the target force is 50N, the actual force is controlled within the range of 45-55N. The switching instruction and collaborative flag generated by the advanced response layer are timestamped by the hardware timestamp synchronization engine and synchronized to the operating mechanism controller and chassis controller via industrial Ethernet. If the time when the operating mechanism completes the mode switch and the time when the chassis receives the flag deviates by more than 2ms, the system automatically inserts an empty instruction cycle until the timing is aligned to ensure the temporal and spatial consistency of force control and chassis compensation, and avoid compensation lag or overshoot caused by delays.

[0047] Through the above three-level response layer and dynamic switching mechanism, the system realizes the intelligent processing of the entire process from contact force analysis, sliding probability prediction to control mode switching, while ensuring the precise coordination of chassis compensation and operating mechanism control, significantly improving the operational stability and reliability of humanoid robots in scenarios such as precision assembly and flexible grasping.

[0048] During the dual-system collaborative compensation process, to resolve the timing deviation problem between the operating mechanism and the chassis, the system uses a time-space stamp synchronization engine to build a high-precision time alignment mechanism. The specific implementation is as follows:

[0049] Between operating mechanism control and chassis motion control:

[0050] After the tactile signals and chassis posture data are collected by the operating mechanism controller and chassis controller respectively, they are time-stamped by the hardware-level time and space stamp synchronization engine. The main controller calculates the timestamp deviation between the tactile signals and chassis posture data in real time and dynamically compensates when the timestamp deviation exceeds 1ms:

[0051] If the tactile signal precedes the chassis data, an interpolation algorithm is used to estimate the chassis posture change within the timestamp deviation period. If the chassis data is ahead, the tactile signal is delayed and buffered until the time deviation between the two is less than 1ms. For example, when the timestamp deviation is 1.5ms, the system estimates the translation within the timestamp deviation to be 0.15mm based on the chassis's motion speed. This estimate is embedded in the tactile signal processing flow, reducing the data synchronization error to within ±0.5ms. The operating mechanism control instructions and chassis execution instructions are transmitted through a shared memory queue. The main controller monitors the time difference between instruction transmission and execution in real time. If the deviation between the operating mechanism control instruction and chassis execution is greater than 2ms, an idle instruction cycle of 2ms is automatically inserted to suspend the transmission of new instructions until the previous instruction is executed and the status is fed back. For example, if the operating mechanism sends a "switch to force tracking control" instruction, if the chassis's inertia causes an execution delay of 3ms, the system inserts an idle instruction cycle to restore the timing deviation to within 1ms, ensuring that force control adjustment and chassis compensation are synchronized and effective.

[0052] The clocks of each controller are regularly calibrated through a distributed clock synchronization protocol, and the time drift within the daily calibration interval is less than 1μs. At the same time, a historical database of timing deviations is established, and machine learning algorithms are used to predict timing delays under different working conditions, and compensation parameters are adjusted in advance. This effectively solves the problem of action lag or advance caused by clock asynchrony in traditional collaborative control, improves the collaborative efficiency of the operating mechanism and chassis, and enhances the system's spatiotemporal consistency and control accuracy.

[0053] During the chassis' reverse translation compensation process, to avoid attitude instability caused by reaction force, the system uses the inertial measurement unit to monitor dynamic disturbances in real time and generate correction components. The specific implementation is as follows:

[0054] A three-axis accelerometer and a three-axis gyroscope are integrated at the bottom of the chassis to collect acceleration and angular velocity signals of the chassis at a frequency of 500Hz when compensating for translation. For example, when the chassis translates along the negative direction of the x-axis to compensate for the end-position drift, the accelerometer simultaneously monitors the linear acceleration change caused by the x-axis reaction force, and the gyroscope tracks the angular acceleration around the z-axis, that is, the attitude oscillation amplitude, to ensure comprehensive perception of the six-degree-of-freedom motion state. The system presets the attitude oscillation amplitude threshold to be , the corresponding gyroscope output angular velocity is greater than 1° / s. When the angular velocity around any axis exceeds the threshold for three consecutive sampling periods, it is determined that there is a risk of secondary disturbance and the reverse compensation mechanism is triggered immediately. For example, if the gyroscope detects an angular velocity of 1.2° / s around the z-axis, it indicates that the chassis is rotating and oscillating due to the reaction force, and a correction component needs to be generated to offset the disturbance. The reverse compensation vector is calculated based on the inertial measurement unit data through the proportional-differential controller. For angular velocity disturbances, an angular acceleration command opposite to the oscillation direction is generated. For example, when oscillating in the positive direction around the z-axis, an angular acceleration of -0.1° / s² is output. For linear acceleration disturbances, the chassis translation speed is adjusted, such as reducing the x-axis translation speed from 0.1m / s to 0.08m / s. The correction component is injected in real time through the chassis motion controller with a response delay of less than 10ms, so that the attitude oscillation amplitude is decayed to below the threshold within 50ms. For example, in a certain compensation, the chassis is generated due to the reaction force. After the reverse compensation vector is executed, the oscillation amplitude is reduced to , effectively suppressing secondary disturbances. To improve the correction accuracy, the position data of the lidar and the inertial measurement unit signal are integrated, and the true posture of the chassis is estimated by the extended Kalman filter algorithm to avoid the noise interference of a single sensor. At the same time, the PD controller gain is dynamically adjusted according to the compensation translation speed. When the speed is greater than 0.2m / s, the gain coefficient is increased to 1.2 to enhance the response sensitivity. When the speed is less than or equal to 0.2m / s, the gain coefficient is maintained at 1.0 to ensure stability. It is measured that this mechanism can improve the efficiency of posture oscillation suppression in high-speed compensation scenarios. Compared with the traditional open-loop compensation scheme, it reduces the occurrence rate of secondary disturbances and significantly improves the stability of the compensation process and the overall motion safety of the robot.

[0055] When the end of the operating mechanism contacts the workpiece, in order to predict the risk of sliding and suppress the accumulation of kinetic energy, the system establishes a sliding energy integration model to accurately determine the critical sliding stage. The specific implementation method is as follows:

[0056] The determination of critical slip stage includes:

[0057] A sliding energy integration model is established. Based on the contact mechanics theory, the tangential acceleration signal of the contact surface at the end of the operation is converted into As the core monitoring parameter, the accumulated sliding energy is obtained through integral operation. , the calculation formula is ;in is the equivalent mass of the contact object, The tangential velocity is collected synchronously by the force sensor and the acceleration sensor. and tangential acceleration ,use Dynamic solution of equivalent mass , to ensure the real-time performance of model parameters, preset the material friction coefficient database, store the critical sliding energy thresholds of different material combinations, and when the workpiece material is detected to be steel, the system automatically calls the corresponding threshold , and combined with the current contact area correction threshold to avoid misjudgment caused by contact area differences, when the sliding energy integral value exceeds the material friction coefficient-related threshold, the system triggers the critical sliding response 50 milliseconds in advance, and uses the force control module to linearly reduce the end-operation contact force by 20% from the current value to suppress the accumulation of sliding kinetic energy. At the same time, a pre-compensation instruction is sent to the chassis to reduce the positive pressure of the contact surface and adjust the chassis posture in advance, thereby suppressing the further development of the sliding trend. The relative displacement of the contact surface is monitored in real time through the visual sensor. If micro-slip still occurs after the early response is detected, that is, the displacement is greater than 0.1mm, the contact force reduction amplitude is automatically increased, and the chassis pre-compensation time is extended to 100 milliseconds, thereby improving the operational stability and safety of the robot in different material contact scenarios.

[0058] When the chassis performs translational compensation motion, to avoid interference with the manipulator's motion trajectory and ensure smooth motion, the system builds an obstacle avoidance model based on the manipulator's kinematic chain and optimizes the compensation trajectory. The specific implementation is as follows:

[0059] When the chassis translation motion is executed, an obstacle avoidance interference domain is constructed based on the operating arm kinematic chain. First, the operating arm kinematics forward solution algorithm is used to calculate the end position of the manipulator arm and the spatial envelope range of the connecting rod corresponding to each joint angle in real time, and a three-dimensional obstacle avoidance interference domain including the base, upper arm, lower arm and other components is constructed. For example, when the upper arm of the manipulator arm rotates to a horizontal position around the shoulder joint, its spatial envelope domain is a fan-shaped area with a radius of 0.8 meters. The chassis compensation path needs to automatically avoid this area. The distance between the chassis planning path and the interference domain is judged in real time through the collision detection algorithm. When the distance is less than 0.2 meters, the path re-planning is triggered to ensure that the safety distance between the two is greater than or equal to 0.3 meters, avoiding the risk of collision caused by the blind spot of the manipulator movement, and automatically avoiding the restricted area of ​​the manipulator arm workspace in the compensation path. In order to control the chassis acceleration to be continuous and without mutation, a cubic B-spline curve is used to fit the compensation path, with the initial position as the starting point and the target compensation path as the target. The system uses the position as the end point and inserts at least three intermediate control points. The curve curvature is optimized by adjusting the control point weight coefficients. For example, when the chassis needs to translate 0.5 meters along the negative x-axis while avoiding the interference zone of the robot arm along the y-axis, a B-spline curve with two turning points is planned to uniformly accelerate the chassis translation speed from 0 to 0.1 m / s and then decelerate uniformly to 0. The peak acceleration is controlled within 0.3 m / s² to avoid the start-stop shock caused by traditional linear interpolation. Physical constraints are applied based on the chassis kinematic model to ensure that the compensation trajectory meets the chassis turning radius limit. The chassis wheel speed is monitored in real time using encoders. If a wheel speed is detected exceeding the rated value, the curve curvature is automatically adjusted to reduce the load on that wheel. For example, when the left front wheel speed reaches 350 rpm, the system offsets the steering control point to the right by 0.1 meter, reducing the left and right wheel speed difference by 20 rpm and ensuring that the chassis movement is within the allowable range of mechanical performance.

[0060] Through the above-mentioned obstacle avoidance and trajectory optimization mechanism, the system realizes intelligent avoidance of the chassis compensation path and the robotic arm workspace. The significantly improved motion smoothness effectively extends the service life of the robot's key components and provides safe and reliable motion control guarantees for multi-degree-of-freedom collaborative operations.

[0061] In the dual-system collaborative compensation process, in order to resolve the conflict between the operating mechanism and the chassis control command, the system sets a collaborative arbitrator to achieve dynamic priority adjustment. The specific implementation method is as follows:

[0062] A collaborative arbitrator is set up. When the impedance control output of the operating mechanism conflicts with the chassis compensation vector, the control weight is reallocated with the goal of minimizing the end-point positioning error. The collaborative arbitrator monitors the direction and amplitude differences between the impedance control output of the operating mechanism and the chassis compensation vector in real time. When the two commands conflict on the same degree of freedom, such as the operating mechanism requires a 0.1mm movement along the positive X-axis and the chassis compensation vector requires a 0.2mm movement along the negative X-axis, the control weight is reallocated using a quadratic programming algorithm with the optimization goal of minimizing the end-point positioning error in the work coordinate system. For example, the influence weights of the two control commands on the end-point positioning error are calculated, and a fused composite control command is generated to ensure that the error is lower than before the conflict. A conflict arbitration strategy is then formulated to prioritize the chassis compensation vector. The preset arbitration priority rule is "chassis compensation vector first." When an irreconcilable command conflict is detected, such as when the operating mechanism force control and the chassis translation direction are completely opposite, the arbitrator immediately freezes the active control output of the operating mechanism, forcing it to switch to passive compliance mode and fully execute the chassis compensation vector. For example, when the operating mechanism attempts to increase the clamping force due to workpiece adhesion, and the chassis When reverse translation is required to avoid a collision, the arbiter prioritizes chassis translation commands, while the manipulator passively follows the displacement to avoid mechanical overload caused by deadlock. The manipulator then switches to passive compliance mode. Once in passive compliance mode, torque sensors monitor joint loads in real time. If the torque of a joint exceeds 80% of its rated value, an emergency stop logic is immediately triggered, disabling chassis compensation vector output and initiating the arm's retraction sequence. Simultaneously, tactile sensors monitor changes in contact force. When a sudden drop in contact force exceeds 50%, the arbiter determines a risk of workpiece fall and automatically terminates all motion commands, ensuring operational safety. This collaborative arbitration mechanism effectively addresses the positioning deviation and mechanical damage caused by command conflicts in traditional dual-system control, improves the control accuracy of end-positioning errors in conflicting scenarios, and reduces the failure rate of manipulator joint overload. In practical applications, this mechanism can quickly resolve conflicts between autonomous robot decisions and manual intervention commands in human-robot collaborative assembly scenarios, ensuring the safety of human-robot collaboration while ensuring control command consistency under complex working conditions. This significantly improves the reliability and robustness of multi-system collaborative operations of humanoid robots.

[0063] During force tracking impedance control, to adapt to the contact characteristics of objects with different stiffnesses and optimize compensation, the system implements intelligent control through tactile signal spectrum analysis and virtual coordinate system rendering technology. The specific implementation is as follows:

[0064] The operations when force tracking impedance control is executed are as follows:

[0065] The stiffness characteristics of the contact object are identified through high-frequency spectrum analysis of the tactile signal, and the force tracking gain is dynamically adjusted. Objects that withstand a force of more than 100 Newtons per millimeter of length are recorded as high-stiffness objects, and low-gain slow tracking is used for high-stiffness objects. Objects that withstand a force of less than 30 Newtons per millimeter of length are recorded as low-stiffness objects, and high-gain fast tracking is used for low-stiffness objects.

[0066] The tactile signal is subjected to a high-frequency spectrum analysis of 100Hz to 1000Hz by Fourier transform, the signal energy distribution characteristics are extracted, and the contact force per unit length is calculated, that is, the force value / contact area. If the force per millimeter of length exceeds 100 Newtons, it is judged as a high-rigidity object. At this time, the force tracking gain coefficient is reduced from the default value of 1.0 to 0.3, reducing the force control response speed and avoiding the impact load caused by rigid contact. If the force per millimeter of length is less than 30 Newtons, it is judged as a low-rigidity object. The gain coefficient is increased to 1.5, and the force tracking speed is accelerated to maintain stable contact. For example, when contacting metal gears, low gain control makes the contact force fluctuation amplitude change from ±20 The error rate is reduced to ±6 Newtons, which reduces the damage rate of parts. When contacting the rubber sealing ring, high-gain control reduces the force tracking error and improves assembly efficiency. In the virtual coordinate system at the end of the operating mechanism, the compensation trajectory is rendered in real time with a resolution of 0.1 mm, and a posture error heat map is generated through a color gradient, where red represents high error and blue represents low error. When 5 consecutive pixels in a certain area in the heat map display red, it is judged as local error accumulation. The system automatically increases the compensation vector weight of the area from the default 1.0 to 1.8 to enhance the compensation strength. The Gaussian kernel function is used to smooth the edges of the heat map to avoid control oscillations caused by weight mutations and ensure a smooth and continuous compensation process.

[0067] Through the above-mentioned stiffness adaptive control and error heat map optimization mechanism, the system realizes refined force control of objects of different materials. The compliance of the operating mechanism when in contact with flexible objects is significantly enhanced, effectively solving the problem of insufficient adaptability of traditional fixed gain control in multi-material operations, and providing key technical support for the widespread application of humanoid robots in scenarios such as precision manufacturing and service robots.

[0068] S4. Use the deep deterministic policy gradient algorithm to optimize the mapping relationship between the sliding trend judgment threshold and the chassis compensation vector amplitude coefficient.

[0069] The deep deterministic policy gradient algorithm optimization process includes:

[0070] A compensation vector-task scenario association knowledge base is constructed to store the historical optimal compensation parameters of different operation tasks. When the similarity between the detected task scenario and the historical task scenarios in the compensation vector-task scenario association knowledge base reaches a similarity threshold, the compensation vector is loaded with the parameters from the compensation vector-task scenario association knowledge base to initialize the compensation vector.

[0071] When optimizing the mapping relationship between the sliding trend determination threshold and the chassis compensation vector, to improve the algorithm convergence speed and the scenario adaptability of the compensation parameters, the system uses a deep deterministic policy gradient algorithm combined with historical data to build an intelligent optimization mechanism. The specific implementation method is as follows:

[0072] The key feature parameters of different operation tasks are collected through sensors, and the historical optimal compensation parameters are recorded synchronously to form a structured knowledge base. When a new task is triggered, the system extracts the feature parameters of the current scene, such as identifying the color of the object and judging the material through the texture through the visual sensor, obtaining the operation speed through the encoder, and using the cosine similarity algorithm to calculate the matching degree with the historical tasks in the knowledge base. The similarity threshold is set to 0.8. When the matching degree is greater than or equal to 0.8, the corresponding historical parameter group is automatically loaded to initialize the compensation vector to avoid the time-consuming problem caused by training from scratch. The historical parameters are used as the initial values, and the strategy function is constructed through the deep neural network. The input is the scene feature vector and the output is the optimized compensation vector. The compensation parameter vector is stored in the experience replay mechanism to store historical interaction data. The reward function is defined as the inverse of the terminal positioning error. The network parameters are updated through the temporal difference algorithm to make the policy function gradually approach the optimal solution. When the scene characteristics suddenly change during the task execution, the system calculates the error value between the current compensation effect and the historical parameters in real time. If the error is greater than 15%, the online learning mode is automatically triggered, and the policy function parameters are fine-tuned using the current real-time data to ensure that the compensation parameters dynamically adapt to the scene changes. In the material mixed scene, this mechanism makes the adaptive adjustment time of the compensation parameters less than 200 milliseconds, reduces the fluctuation amplitude of the positioning error, and enhances the adaptability of the humanoid robot to unstructured environments.

[0073] The present invention uses multi-dimensional tactile sensors to monitor contact force vectors and sliding trends in real time, maps posture drift to the chassis coordinate system based on the spatiotemporal synchronization layer of Lie group constraints, generates compensation vectors, and adopts a three-level tactile response layer to dynamically switch the force / position mixed impedance mode. The drift is offset by the reverse translation of the omnidirectional chassis, and the compensation parameters are optimized through a deep deterministic policy gradient algorithm. The system ensures the timing alignment of control instructions through a spatiotemporal stamp synchronization engine, uses an inertial measurement unit to suppress posture oscillation during the compensation process, constructs an obstacle avoidance interference domain based on the kinematic chain of the manipulator arm, and uses a B-spline curve to smooth the compensation trajectory, thereby improving the accuracy, stability and safety of the humanoid robot in complex contact operations, and providing an efficient and reliable solution for human-machine collaboration scenarios.

[0074] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A multimodal environment perception and adaptive chassis control method for a humanoid robot, characterized in that: The following steps are involved: S1. Use a multi-dimensional tactile sensor to monitor the contact force vector and sliding trend of the end of the operating mechanism in real time. When the contact force vector exceeds a preset safety threshold or the axial deviation rate of the sliding trend is greater than 5 mm per second, activate the collaborative compensation mechanism; S2. The collaborative compensation mechanism maps the posture drift of the end of the operating mechanism due to the contact force disturbance to the chassis motion coordinate system based on the spatiotemporal synchronization layer of Lie group constraints, and generates the direction and amplitude of the chassis compensation vector; S3. Dual-system collaborative compensation is performed based on the chassis compensation vector. Dual-system collaborative compensation includes operating mechanism control and chassis motion control. The operating mechanism control dynamically switches the force / position hybrid impedance mode according to the sliding trend. That is, position-dominant impedance control is adopted in the stable contact phase, and force-tracking impedance control is switched in the critical sliding phase. The chassis motion control omnidirectionally moves the chassis in the direction opposite to the operating mechanism movement direction, offsetting the posture drift in real time and maintaining the positioning error of the end in the working coordinate system. S4. Use the deep deterministic policy gradient algorithm to optimize the mapping relationship between the sliding trend judgment threshold and the chassis compensation vector amplitude coefficient.

2. The method for multimodal environment perception and adaptive chassis control of a humanoid robot according to claim 1, characterized in that: The sliding trend dynamic switching force / position mixed impedance mode includes: Construct a three-level tactile response layer, where the primary response layer analyzes the contact force vector direction, the intermediate response layer calculates the sliding trend probability distribution, and the advanced response layer combines the two to generate impedance switching instructions; When the sliding tendency probability is greater than the preset threshold, the control switches from position-dominant impedance control to force-tracking impedance control, and the chassis compensation collaborative flag is activated.

3. The method for multimodal environment perception and adaptive chassis control of a humanoid robot according to claim 1, characterized in that: Between the operating mechanism control and the chassis motion control: The tactile signal and chassis posture data are timestamped through the time-space stamp synchronization engine, and dynamic compensation is performed based on the timestamp deviation. If the deviation between the operating mechanism control instruction and the chassis execution is greater than 2 milliseconds, an empty instruction cycle is inserted to align the timing.

4. The method for multimodal environmental perception and adaptive chassis control of a humanoid robot according to claim 1, characterized in that: The chassis is monitored in real time to compensate for the reaction force when it moves in the opposite direction, and the attitude oscillation amplitude during the compensation process is fed back through the chassis inertial measurement unit. If the attitude oscillation amplitude is greater than Then a reverse compensation vector correction component is generated to offset the secondary disturbance.

5. The method for multimodal environment perception and adaptive chassis control of a humanoid robot according to claim 1, characterized in that: The determination of the critical sliding stage includes: A sliding energy integration model is established to integrate the tangential acceleration of the contact surface of the operating end. When the integral value exceeds the threshold associated with the material friction coefficient, the critical sliding response is triggered 50 milliseconds in advance. When triggered, the contact force of the operating end is reduced to suppress the accumulation of sliding kinetic energy.

6. The method for multimodal environment perception and adaptive chassis control of a humanoid robot according to claim 1, characterized in that: When the chassis translation motion is executed, an obstacle avoidance interference domain is constructed based on the kinematic chain of the manipulator arm, and the restricted area of ​​the manipulator arm workspace is automatically avoided in the compensation path. A B-spline curve is used to smooth the compensation trajectory to control the chassis acceleration to be continuous and without sudden changes.

7. The method for multimodal environment perception and adaptive chassis control of a humanoid robot according to claim 1, characterized in that: A collaborative arbitrator is set up. When the impedance control output of the operating mechanism conflicts with the chassis compensation vector, the control weight is redistributed with the goal of minimizing the end positioning error. A conflict arbitration strategy is formulated to prioritize the chassis compensation vector and switch the operating mechanism to passive compliance mode.

8. The method for multimodal environment perception and adaptive chassis control of a humanoid robot according to claim 1, characterized in that: The operation of the force tracking impedance control is as follows: The stiffness characteristics of the contact object are identified through high-frequency spectrum analysis of the tactile signal, and the force tracking gain is dynamically adjusted. Objects that withstand a force of more than 100 Newtons per millimeter of length are recorded as high-stiffness objects, and low-gain slow tracking is used for high-stiffness objects. Objects that withstand a force of less than 30 Newtons per millimeter of length are recorded as low-stiffness objects, and high-gain fast tracking is used for low-stiffness objects.

9. The method for multimodal environment perception and adaptive chassis control of a humanoid robot according to claim 1, characterized in that: Fourteen compensation trajectories are rendered in the virtual coordinate system of the end of the operating mechanism to form a posture error heat map, where red represents high error and blue represents low error. When five consecutive pixels in a certain area of ​​the heat map display red, it is determined to be local error accumulation. When a certain area of ​​the posture error heat map displays local error accumulation, the compensation vector weight of the area is automatically enhanced.

10. The method for multimodal environment perception and adaptive chassis control of a humanoid robot according to claim 1, characterized in that: The deep deterministic policy gradient algorithm optimization process includes: A compensation vector-task scenario association knowledge base is constructed to store the historical optimal compensation parameters of different operation tasks. When the similarity between the detected task scenario and the historical task scenarios in the compensation vector-task scenario association knowledge base reaches a similarity threshold, the compensation vector is loaded with the parameters from the compensation vector-task scenario association knowledge base to initialize the compensation vector.

Images