A six-dimensional force space decoupling method, system, device and storage medium based on a humanoid robot

Abstract

The application relates to a six-dimensional force sensation space decoupling method, system, device and storage medium based on a humanoid robot, and relates to the technical field of humanoid robots. The six-dimensional force sensation space decoupling method comprises the following steps: synchronously collecting and cleaning multi-dimensional strain signals of the robot to obtain effective moment data of different dimensions; nonlinearly coupling and expanding the effective moment data to establish a static six-dimensional force space and calculate a static local moment; according to joint kinematics and a robot joint motion trajectory, a six-dimensional inertia vector is constructed to compensate the static local moment and obtain a dynamic local moment; the dynamic local moment is decoupled according to a discrete temperature field sequence, zero space drift is inhibited, and a stable local moment is obtained; the stable local moment is projected and transformed to a robot centroid coordinate system, and global force sensation semantics are output; a static six-dimensional force space is constructed through nonlinear coupling, joint kinematics inertia compensation is combined, and global force sensation semantics are output, so that the multi-dimensional force perception accuracy of the robot is improved.

Description

Technical Field

[0001] This application relates to the field of humanoid robot technology, and in particular to a method, system, device and storage medium for six-dimensional force perception space decoupling based on humanoid robots. Background Technology

[0002] Humanoid robots, as the core physical carriers of embodied intelligence, possess a fundamental advantage in seamlessly integrating into the social and production environments constructed for humans through their human-like form—bipedal walking and dexterous hand manipulation. Through real-time interaction with physical entities, they achieve an integrated closed loop of perception, cognition, decision-making, and action. Within this loop, tactile perception, especially six-dimensional force sensing, is a cornerstone technology for achieving precise manipulation, stable walking, and safe human-robot collaboration. Six-dimensional force sensors can simultaneously measure forces (Fx, Fy, Fz) and torques (Mx, My, Mz) in three-dimensional space, endowing robots with realistic mechanical perception capabilities in their wrists, ankles, and fingertips. This allows them to perceive contact forces, gravity compensation, and external pushes and pulls, just like humans, providing essential feedback for motion control and compliant interaction.

[0003] The cutting-edge technologies surrounding the development of humanoid robots are currently focusing on three major dimensions: perception and decision-making, motion control, and energy and hardware. Six-dimensional force sensors play a crucial guiding role across these areas. In industrial scenarios, they enable humanoid robots to precisely control contact forces in precision assembly in automobile manufacturing, insertion and removal of electronic components, and compliant installation of photovoltaic panels, preventing workpiece damage and replacing traditional fixed-position specialized machines, truly unleashing flexible manufacturing capabilities. In service scenarios, six-dimensional force sensing allows rehabilitation training robots to perceive the subtle intentions of patients, provides surgical assistance robots with delicate tissue tactile sensation, and enables family companion robots to apply appropriate force when helping or passing items, ensuring absolute safety. In specialized operations, it enables humanoid robots to achieve adaptive contact with unknown rigid environments through force-position mismatch in high-risk tasks such as valve operation in nuclear power plants, rubble demolition, and on-orbit maintenance on space stations, avoiding secondary accidents caused by force-position mismatch.

[0004] A robot experimental platform and operating method based on a serial six-dimensional force sensor are disclosed. The platform includes a fixed horizontal experimental platform with an adjustable base mounted below it. A mounting panel is located above the platform for mounting the robot and other accessories. The platform features a robot base mounted on its surface, a robot arm mounted on the base, and a TCP disk attached to the robot arm for connecting to a flange. A six-dimensional force sensor is mounted through the flange. The method for using this platform is also disclosed. This invention solves the problem of current experimental operating platforms being cumbersome and difficult to adapt to different assembly scenarios. It significantly improves the robot's working efficiency, enabling it to complete complex tasks and greatly enhancing its assembly accuracy.

[0005] The existing technical solutions mentioned above have the following drawbacks: 1. The joint motors and actuators of humanoid robots generate local high temperatures during operation, and the thermal sensitivity of strain gauges and semiconductor sensitive elements leads to significant temperature drift, while achieving online dynamic compensation greatly increases the complexity of the system; 2. The limbs of humanoid robots require extremely high rigidity to ensure control bandwidth and positioning accuracy, making it difficult to simultaneously achieve strong impact resistance and sensitive detection of small forces (such as threading a needle), which limits the application in all scenarios. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this application is to provide a six-dimensional force perception space decoupling method, system, device, and storage medium based on humanoid robots. By synchronously acquiring and cleaning strain signals through multi-channel Kalman filtering, a static six-dimensional force space is constructed using nonlinear coupling. Combined with joint kinematic inertial compensation and temperature drift suppression, the system is projected and transformed to the centroid coordinate system to output global force perception semantics, which significantly improves the accuracy of multi-dimensional force perception and dynamic anti-interference capability of the robot.

[0007] The objective of this application is achieved through the following technical solution: Firstly, this application provides a six-dimensional force perception space decoupling method based on a humanoid robot, including: By synchronously collecting and cleaning the robot's multidimensional strain signals through multidimensional data channels, effective torque data in different dimensions are obtained. Nonlinear coupling and extended effective torque data are used to establish a static six-dimensional force space and calculate static local torques. Based on joint kinematics and the robot's joint motion trajectory, a six-dimensional inertial vector is constructed, and the static local torque is compensated to obtain the dynamic local torque; By decoupling the dynamic local torque based on the discrete temperature field sequence and suppressing the zero-point spatial drift, a stable local torque is obtained. Project and transform the stable local torque to the robot's center of mass coordinate system, and output the global force perception semantics.

[0008] By adopting the above technical solution, strain signals are collected synchronously through multiple channels and cleaned by Kalman filtering. A static six-dimensional force space is constructed using nonlinear coupling. Combined with joint kinematic inertial compensation and temperature drift suppression, the signal is projected and transformed to the centroid coordinate system to output global force perception semantics, which significantly improves the robot's multi-dimensional force perception accuracy and dynamic anti-interference capability.

[0009] This application is further configured to: synchronously acquire and clean the multidimensional strain signals of the robot based on multidimensional data channels to obtain effective torque data in different dimensions, including: Time alignment is performed on the data transmission channel, and a multi-dimensional data channel is constructed by combining data sampling points; Based on the preset data acquisition frequency and multi-dimensional data channels, the joint strain state of the robot is monitored and raw strain signals are collected. The original strain signal is filtered to remove noise and artifacts, resulting in a noise-free strain signal. Peak detection is performed on the noiseless strain signal, the differential amplitude is calculated, and it is compared with the preset standard threshold to determine the peak point and the effective point. If the differential amplitude is greater than the standard threshold, the current data sampling point is determined to be a peak point, and it is filled according to the linear interpolation of adjacent valid points; If the amplitude of the effective point is greater than the preset full-scale value, the current effective point is determined to be overloaded and marked as an invalid impact, thus obtaining an effective strain signal; Temperature monitoring is performed at data sampling points to collect temperature data and form a local temperature gradient sequence; Based on a lightweight fully connected network, temperature drift inference is performed on the local temperature gradient sequence and running timestamps to calculate the zero-point drift. The effective strain signal is zero-point corrected and analog-to-digital converted based on the zero-point drift to obtain the steady-state strain voltage value. By extending and decoupling the steady-state strain voltage value, effective torque data in different dimensions are obtained.

[0010] By adopting the above technical solution, multi-dimensional channels are constructed through time alignment, peaks are detected and filled by differential thresholding and linear interpolation, and temperature drift inference and zero-point correction are achieved by combining lightweight fully connected network, dynamic filtering and high-precision decoupling of strain signals are realized, which significantly improves the noise resistance and temperature stability of multi-dimensional torque data.

[0011] This application further includes: nonlinear coupling and extension of effective torque data, establishing a static six-dimensional force space, and calculating static local torques, including: The effective torque data is normalized based on the full-scale value to obtain the standard torque data; Dimensional expansion is performed on the standard torque data to construct an expanded feature vector containing linear terms, square terms, and cross-product terms; Based on the mapping reasoning of the extended eigenvectors using the nonlinear decoupling matrix, a static six-dimensional force space is generated. Based on the preset local coordinate system reference, the static six-dimensional force space is vector-combined to calculate the static local torque.

[0012] By adopting the above technical solution, an eigenvector containing nonlinear terms is constructed through normalization and interdimensional expansion, and a static six-dimensional force space is generated by using decoupling matrix mapping. This achieves high-precision decoupling calculation of local torque, significantly improving the accuracy and anti-interference capability of multidimensional force measurement.

[0013] This application further specifies: based on joint kinematics and the robot's joint motion trajectory, a six-dimensional inertial vector is constructed, and static local torques are compensated to obtain dynamic local torques, including: The motion trajectory of the robot's joints is calculated based on joint kinematics to obtain a set of motion parameters; The acceleration of the centroid of the sampling point is calculated based on the position vector of the centroid of the sampling point and the linear acceleration, angular velocity and angular acceleration in the set of motion parameters. Based on the centroid acceleration at the sampling point and the local gravitational components in the motion parameter set, the inertial force and inertial torque are calculated, and a six-dimensional inertial vector is synthesized. The original dynamic torque is obtained by performing compensation calculations on the static local torque and the six-dimensional inertia vector. The original dynamic torque is filtered for noise reduction and peak identification to screen out dynamic local torques.

[0014] By adopting the above technical solution, motion parameters are calculated based on joint kinematics, and a six-dimensional inertial vector is synthesized by combining the center of mass acceleration and gravity components. Static torque is compensated and filtered, and high-precision real-time extraction of dynamic local torque is achieved, which significantly improves the accuracy and response speed of the robot's dynamic force perception.

[0015] This application further specifies: decoupling the dynamic local torque based on the discrete temperature field sequence and suppressing zero-point spatial drift to obtain a stable local torque, including: The data sampling points are synchronously measured based on the global sampling timestamp to generate a discrete temperature field sequence and extract the extended temperature feature vector. Based on the temperature drift torque coupling model, the extended temperature eigenvector is inferred, and the six-dimensional force drift estimate is calculated. Decoupling compensation is performed on the dynamic local torque based on the six-dimensional force drift estimate to obtain the contact torque magnitude; If the robot joint is stationary and the contact torque magnitude is less than the preset silence threshold, the robot's motion state is determined to be a silent period without external force, and the current contact torque magnitude is subjected to an exponential moving average to extract the residual zero-point offset value. If the residual zero offset value remains stable and non-zero for N periods of no external force, it is added to the drift compensation amount as a new zero correction term to update the global zero offset. If the temperature difference between data sampling points in adjacent sampling periods is greater than the preset temperature threshold, the robot's motion state is determined to be a temperature shock period, and the compensation weight of the current contact torque modulus is temporarily frozen until the temperature difference converges. After stabilizing the contact torque modulus by both temperature and zero point, a stable local torque is obtained.

[0016] By adopting the above technical solution, the six-dimensional force drift is compensated based on the temperature drift torque coupling model. The zero-point bias is updated by exponential averaging during the no-external-force quiescent period, and the weights are frozen during the temperature shock period. This achieves temperature decoupling and zero-point drift suppression of dynamic local torque, and improves the stability and robustness of torque data.

[0017] This application is further configured to: project and transform stable local torques to the robot's center-of-mass coordinate system, and output global force perception semantics, including: Based on the data dimensions, stable local torques are combined to construct local force spinors in each dimension; The global force spinor is obtained by projecting the local force spinor to the robot's center of mass coordinate system using the adjoint transformation matrix. Based on the data sampling points and the global force rotation, the robot's contact state is determined and multi-point contact configurations are identified. Semantic mapping of stable local torques is performed based on multi-point contact configurations to extract semantic features of force vector direction and transient features of torque. Based on the current action primitives, the transient features of torque are associated with the task stage to generate task action semantics; The semantics of task actions and force vector directions are annotated to obtain global force perception semantics.

[0018] By adopting the above technical solution, the local force spinor is projected onto the robot's center of mass coordinate system based on the adjoint transformation matrix. Through contact state determination and multi-point contact configuration recognition, combined with action primitive association, the semantic annotation of force vector direction and task action is realized, which significantly improves the globality, semantic level and adaptability of robot force perception and complex interactive tasks.

[0019] Secondly, this application also provides a six-dimensional force perception space decoupling system based on a humanoid robot, which adopts the following technical solution: A six-dimensional force perception space decoupling system based on a humanoid robot, comprising a method for implementing six-dimensional force perception space decoupling, including: The data acquisition and cleaning module is used to synchronously acquire and clean the robot's multidimensional strain signals according to the multidimensional data channel, and obtain effective torque data in different dimensions. The static coupling module is used for nonlinear coupling and expanding effective torque data, establishing a static six-dimensional force space, and calculating static local torques. The dynamic compensation module is used to compensate for static local torques and obtain dynamic local torques based on joint kinematics and robot joint motion trajectories. The temperature drift compensation module is used to decouple dynamic local torques based on discrete temperature field sequences and suppress zero-point spatial drift to obtain stable local torques. The decoupling transformation module is used to project and transform stable local torques to the robot's center-of-mass coordinate system and output global force semantics.

[0020] By adopting the above technical solution, a static six-dimensional force space is constructed through multi-channel synchronous acquisition and nonlinear coupling. Combined with joint kinematic inertial compensation and temperature drift suppression, the force space is projected onto the center of mass to output global force perception semantics, thereby improving the robot's multi-dimensional force perception accuracy, dynamic anti-interference capability, and task semantic level.

[0021] Thirdly, this application also provides an electronic device, comprising: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by one or more processors, the one or more processors implement any of the methods in the above scheme.

[0022] Fourthly, this application also provides a storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the six-dimensional force perception space decoupling method based on a humanoid robot as described above.

[0023] In summary, the beneficial technical effects of this application are as follows: By constructing a static six-dimensional force space through multi-channel synchronous acquisition and nonlinear coupling, and combining joint kinematic inertial compensation and temperature drift suppression, the force space is projected onto the center of mass to output global force perception semantics, thereby improving the robot's multi-dimensional force perception accuracy, dynamic anti-interference capability and task semantic level. Based on the temperature drift torque coupling model to compensate for six-dimensional force drift, the zero-point bias is updated by exponential averaging during the no-external-force quiescent period, and the weights are frozen during the temperature shock period. This achieves temperature decoupling and zero-point drift suppression of dynamic local torque, improving the stability and robustness of torque data. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the six-dimensional force perception space decoupling method in this application; Figure 2 This is a flowchart illustrating step S4 in this application; Figure 3 This is a schematic diagram of the six-dimensional force perception space decoupling system in this application. Detailed Implementation

[0025] The present application will be further described in detail below with reference to the accompanying drawings.

[0026] Firstly, referring to Figure 1 This application discloses a six-dimensional force perception space decoupling method based on a humanoid robot, comprising: S1: Based on the multidimensional data channel, the robot's multidimensional strain signal is synchronously collected and cleaned to obtain effective torque data in different dimensions; S2: Nonlinear coupling and extended effective torque data are used to establish a static six-dimensional force space and calculate static local torques; S3: Based on joint kinematics and the robot's joint motion trajectory, a six-dimensional inertial vector is constructed, and the static local torque is compensated to obtain the dynamic local torque; S4: Decouple the dynamic local torque based on the discrete temperature field sequence and suppress zero-point spatial drift to obtain a stable local torque; S5: Project and transform the stable local torque to the robot's center of mass coordinate system, and output the global force perception semantics.

[0027] In this embodiment, during the delicate maintenance work performed by the humanoid robot in the complex pipelines of a nuclear power plant, the system uses a six-dimensional force sensor array distributed at its ankle, knee, and wrist joints to synchronously collect multi-dimensional strain signals of each joint under bending, torsion, and lateral forces at a sampling rate of 1kHz. After wavelet threshold denoising and Kalman filtering-based multi-channel data alignment and cleaning, the effective torque data of each joint in the X, Y, and Z axes and rotation around the axis are obtained. Subsequently, polynomial basis functions are used to nonlinearly couple and extend the torque to establish a static six-dimensional force field in the six-dimensional tensor space and calculate the static local torque.

[0028] Based on the robot's joint kinematics model and real-time planned foot motion trajectory, a six-dimensional inertial vector containing Coriolis force, centrifugal force, and gravity terms is constructed. This vector is then used as feedforward compensation and superimposed on the static local torque to obtain the dynamic local torque. Furthermore, a discrete temperature field sequence is obtained using a thermocouple array on the joint shell. Principal component analysis is used to decouple the influence of temperature on the zero-point drift of the strain gauges, and a zero-drift suppression filter is employed to obtain the stable local torque. Finally, the stable torque is projected onto the robot's centroid coordinate system through homogeneous coordinate transformation. Global force perception semantics are output by fusing multi-joint data from the whole body, enabling the robot to adaptively adjust its gait and grasping force in high-radiation environments, achieving a force closed-loop control accuracy of 0.1 N·m for precision valve operation.

[0029] Preferably, step S1 includes: Time alignment is performed on the data transmission channel, and a multi-dimensional data channel is constructed by combining data sampling points; Based on the preset data acquisition frequency and multi-dimensional data channels, the joint strain state of the robot is monitored and raw strain signals are collected. The original strain signal is filtered to remove noise and artifacts, resulting in a noise-free strain signal. Peak detection is performed on the noiseless strain signal, the differential amplitude is calculated, and it is compared with the preset standard threshold to determine the peak point and the effective point. If the differential amplitude is greater than the standard threshold, the current data sampling point is determined to be a peak point, and it is filled according to the linear interpolation of adjacent valid points; If the amplitude of the effective point is greater than the preset full-scale value, the current effective point is determined to be overloaded and marked as an invalid impact, thus obtaining an effective strain signal; Temperature monitoring is performed at data sampling points to collect temperature data and form a local temperature gradient sequence; Based on a lightweight fully connected network, temperature drift inference is performed on the local temperature gradient sequence and running timestamps to calculate the zero-point drift. The effective strain signal is zero-point corrected and analog-to-digital converted based on the zero-point drift to obtain the steady-state strain voltage value. By extending and decoupling the steady-state strain voltage value, effective torque data in different dimensions are obtained.

[0030] In this embodiment: In the application of exoskeleton walking robot to assist in torque control of user's knee joint, the system first uses FPGA board to perform microsecond-level time alignment of the data transmission channels of the four strain bridge arms of the six-dimensional force sensor of the knee joint, and constructs synchronous multi-dimensional data channels with 2048Hz data sampling points; according to the preset 2kHz acquisition frequency, the strain state of the robot joint during walking, squatting and standing is monitored in real time through these channels, and the original strain bridge voltage signal is collected.

[0031] Subsequently, a fourth-order Butterworth low-pass filter (cutoff frequency 500Hz) was used to remove high-frequency electrical noise, and human electromyography artifacts were filtered out based on an independent component analysis algorithm to obtain a noise-free strain signal. Slope peak detection was performed on this signal, and the difference amplitude between adjacent sampling points was calculated. When the difference amplitude exceeded a preset threshold of 0.5mV, it was identified as a peak point, and cubic spline interpolation of adjacent valid points was used to fill the peak. Simultaneously, when the amplitude of a valid point exceeded the full scale by 5mV, it was identified as an impact overload and marked as invalid, thus obtaining a valid strain signal. Data was collected every 10ms from the sensor housing. The resistance values ​​at three temperature measurement points form a local temperature gradient sequence (ΔT≤2℃). A lightweight fully connected network takes the temperature gradient and the running timestamp (cumulative 200 hours) as input to infer the zero-point drift (maximum ±0.2mV), performs zero-point correction on the effective strain signal, and converts it into a steady-state strain voltage value using a 24-bit ADC. Finally, the voltage value is extended and decoupled into six-dimensional effective torque data (Fx, Fy, Fz, Mx, My, Mz) using a nonlinear least squares method, which is used to calculate the auxiliary torque in real time to help hemiplegic patients achieve a natural gait.

[0032] Preferably, step S2 includes: The effective torque data is normalized based on the full-scale value to obtain the standard torque data; Dimensional expansion is performed on the standard torque data to construct an expanded feature vector containing linear terms, square terms, and cross-product terms; Based on the mapping reasoning of the extended eigenvectors using the nonlinear decoupling matrix, a static six-dimensional force space is generated. Based on the preset local coordinate system reference, the static six-dimensional force space is vector-combined to calculate the static local torque.

[0033] In this embodiment: In the application of fine force sensing control of the humanoid robot wrist, the system first linearly normalizes the effective torque data obtained from strain signal analysis according to the full-scale value of each channel (e.g., torque range ±10 N·m), obtaining standard torque data in the range [-1, 1]; then, it constructs a system containing linear terms (Fx, Fy, Fz), square terms (Fx², Fy², Fz²), and cross-product terms (FxFy, FyFz, ...). The extended feature vector of FzFx is used to fully express the nonlinear coupling characteristics between multidimensional forces of the sensor. The extended feature vector is linearly mapped and inferred using a 6×9-dimensional nonlinear decoupling matrix obtained in advance through least squares identification. The coupled original signals are mathematically transformed into mutually orthogonal static six-dimensional force space, realizing the physical separation of forces / torques in each dimension. Finally, based on the local coordinate system reference fixed to the robot wrist (X-axis pointing towards the palm, Z-axis perpendicular to the palm), the three force components and three torque components in the static six-dimensional force space are vector synthesized to calculate the static local torque (such as the bending torque around the Y-axis) acting on the center of the wrist joint. This provides high-precision force feedback for the robot to grasp fragile objects and ensures that the grasping force is stably controlled within the preset threshold.

[0034] Preferably, step S3 includes: The motion trajectory of the robot's joints is calculated based on joint kinematics to obtain a set of motion parameters; The acceleration of the centroid of the sampling point is calculated based on the position vector of the centroid of the sampling point and the linear acceleration, angular velocity and angular acceleration in the set of motion parameters. Based on the centroid acceleration at the sampling point and the local gravitational components in the motion parameter set, the inertial force and inertial torque are calculated, and a six-dimensional inertial vector is synthesized. The original dynamic torque is obtained by performing compensation calculations on the static local torque and the six-dimensional inertia vector. The original dynamic torque is filtered for noise reduction and peak identification to screen out dynamic local torques.

[0035] In this embodiment: In the application of humanoid robots carrying heavy loads on rugged mountain roads, the system first calculates the angular displacement and angular velocity of the hip, knee, and ankle joints in real time based on joint kinematics to obtain a set of motion parameters including linear acceleration, angular acceleration, and local gravity components of each sampling point; then, combined with the position vector of the robot's center of mass in each sampling period, the acceleration of the center of mass of the sampling point in the base coordinate system is calculated using the Newton-Euler recursive dynamics formula; then, the acceleration of the center of mass is combined with the gravity component in the motion parameters, and multiplied by the mass and moment of inertia of each link to obtain the inertial force and inertial torque acting at the joint, and assembled into a six-dimensional inertial vector containing six components.

[0036] Next, the static local torque calculated by the six-dimensional force sensor and suppressed by temperature drift is subtracted from the inertial vector to compensate for the loss of inertial influence caused by the robot's own motion, thus obtaining the original dynamic torque. Finally, a second-order low-pass filter (cutoff frequency 30Hz) is applied to the original dynamic torque to eliminate high-frequency vibration noise, and an amplitude threshold (±15% relative to the reference value) is set to identify spikes, eliminating abnormal impact spikes caused by road bumps, and filtering out dynamic local torques that reflect real external contact forces (such as the support torque in the vertical direction of the ankle joint), providing accurate force feedback for the robot to achieve dynamic gait balance and compliant foot force control in complex terrains such as gravel and steep slopes.

[0037] Preferably, refer to Figure 2 Step S4 includes: The data sampling points are synchronously measured based on the global sampling timestamp to generate a discrete temperature field sequence and extract the extended temperature feature vector. Based on the temperature drift torque coupling model, the extended temperature eigenvector is inferred, and the six-dimensional force drift estimate is calculated. Decoupling compensation is performed on the dynamic local torque based on the six-dimensional force drift estimate to obtain the contact torque magnitude; If the robot joint is stationary and the contact torque magnitude is less than the preset silence threshold, the robot's motion state is determined to be a silent period without external force, and the current contact torque magnitude is subjected to an exponential moving average to extract the residual zero-point offset value. If the residual zero offset value remains stable and non-zero for N periods of no external force, it is added to the drift compensation amount as a new zero correction term to update the global zero offset. If the temperature difference between data sampling points in adjacent sampling periods is greater than the preset temperature threshold, the robot's motion state is determined to be a temperature shock period, and the compensation weight of the current contact torque modulus is temporarily frozen until the temperature difference converges. After stabilizing the contact torque modulus by both temperature and zero point, a stable local torque is obtained.

[0038] In this embodiment: In the application of a humanoid robot grasping a hot steel billet in a high-temperature casting workshop, the system synchronously measures the temperature at three temperature measurement points distributed on the shell of the six-dimensional force sensor on the robot's wrist according to the global sampling timestamp, generating a discrete temperature field sequence (sampling rate 100Hz), and extracting an extended temperature feature vector from the sequence, including the average temperature, temperature gradient, and heating rate; then, using a temperature drift torque coupling model (six-input nine-output BP neural network) calibrated in the temperature chamber beforehand, the system infers the feature vector and calculates the estimated six-dimensional force drift value of each force / torque in real time (for example, for every 10°C increase in temperature, the Z-axis force drift increases by 2.5N).

[0039] Subtract the drift estimate from the previously obtained dynamic local torque to obtain the contact torque magnitude reflecting the actual external force contact. When the robot joint remains stationary and the contact torque magnitude is less than the preset silence threshold of 0.5 N·m for 300 ms, it is determined to be a no-external-force silence period. The current magnitude is then subjected to an exponential moving average (smoothing factor 0.05) to extract the residual zero-point offset value. If the residual zero-point offset value (e.g., 0.08 N) is stable and non-zero for 5 consecutive silence periods, it is added to the drift compensation amount as a new zero-point correction term to update the global zero-point bias (corrected from 0.02 N·m to 0.10 N·m). When the temperature difference between adjacent sampling points exceeds the threshold of 1.5℃ / s, it is determined to enter the temperature shock period (e.g., the instant the billet just contacts the object). The compensation weight of the current contact torque magnitude is temporarily frozen (keeping the drift value before the shock) to avoid sudden jumps in the compensation amount due to transient overheating or overcooling, until the temperature difference value converges.

[0040] Ultimately, the stable local torque output after temperature and zero-point stabilization is used for robot force-position hybrid control, enabling it to complete billet handling with a force accuracy of ±0.5N even in a temperature difference environment of 400℃, avoiding slippage or crushing.

[0041] Preferably, step S5 includes: Based on the data dimensions, stable local torques are combined to construct local force spinors in each dimension; The global force spinor is obtained by projecting the local force spinor to the robot's center of mass coordinate system using the adjoint transformation matrix. Based on the data sampling points and the global force rotation, the robot's contact state is determined and multi-point contact configurations are identified. Semantic mapping of stable local torques is performed based on multi-point contact configurations to extract semantic features of force vector direction and transient features of torque. Based on the current action primitives, the transient features of torque are associated with the task stage to generate task action semantics; The semantics of task actions and force vector directions are annotated to obtain global force perception semantics.

[0042] In this embodiment, in the application of humanoid robot-assisted neurosurgical microsurgery, the system combines the six-dimensional stable local torques of the wrist joint into local force spinors according to Fx, Fy, Fz, Mx, My, and Mz. These spinors are then projected onto the robot's centroid coordinate system via an adjoint transformation matrix to obtain the global force spinor. Combined with real-time contact state determination using 1kHz sampling points, the system identifies multi-point contact configurations between the instrument and brain tissue, such as "point separation," "slippage," or "pushing." Subsequently, semantic mapping is performed on the stable local torques based on these configurations to extract "horizontal peeling force" and "vertical lifting force." The system generates "equal force vector direction semantics" and transient torque features such as "tremor frequency peak"; then, based on the current "tumor dissection" action primitive, it associates the transient features with the task stages of "contact probing - blunt separation - sharp detachment" to generate task action semantics such as "capsule is about to rupture" and "sudden drop in resistance"; finally, it jointly annotates the task action semantics and force vector direction semantics to output global force perception semantics (such as "the incision is too deep, please withdraw 0.2mm"), thereby guiding the robot to safely complete meningeal dissection under millinewton-level force perception and avoid traction damage to cortical blood vessels.

[0043] Secondly, referring to Figure 3 This application also provides a six-dimensional force perception space decoupling system based on a humanoid robot, used to implement a six-dimensional force perception space decoupling method, including: The data acquisition and cleaning module is used to synchronously acquire and clean the robot's multidimensional strain signals according to the multidimensional data channel, and obtain effective torque data in different dimensions. The static coupling module is used for nonlinear coupling and expanding effective torque data, establishing a static six-dimensional force space, and calculating static local torques. The dynamic compensation module is used to compensate for static local torques and obtain dynamic local torques based on joint kinematics and robot joint motion trajectories. The temperature drift compensation module is used to decouple dynamic local torques based on discrete temperature field sequences and suppress zero-point spatial drift to obtain stable local torques. The decoupling transformation module is used to project and transform stable local torques to the robot's center-of-mass coordinate system and output global force semantics.

[0044] In this embodiment, during the remote control operation of the humanoid robot for flange docking and bolt tightening of deep-sea oil and gas pipelines, the data acquisition and cleaning module first uses a seven-channel six-dimensional force sensor array distributed on the wrists of the robot's two arms and one ankle joint to collect multi-dimensional strain signals of the robot body under water flow disturbance and robotic arm movement in real time at a synchronous sampling rate of 2kHz. After wavelet threshold denoising and median filtering based on sliding window to remove artifact peaks caused by water particle impact, fifteen-dimensional effective torque data of each joint in the X, Y, Z axes and around the axis are obtained.

[0045] The static coupling module uses nonlinear polynomial basis functions containing linear, quadratic, and cross-product terms to extend the dimensions of these torque data. It then establishes a static six-dimensional force space using a 9×15 decoupling matrix identified by least squares, mapping the coupling signals into physically orthogonal six-dimensional force / torque components. Subsequently, it calculates the static local torque based on the vector combination of the local coordinate system of each joint, making the physical meaning of the force on each joint clear. The dynamic compensation module, based on the robot joint kinematic model (including underwater added mass and Coriolis force terms) and preset underwater crawling and bolt tightening trajectories, calculates the linear acceleration, angular velocity, and angular acceleration of the center of mass in real time. It synthesizes a six-dimensional inertial vector containing inertial force, centrifugal force, and gravitational components, and subtracts this inertial vector from the static local torque to obtain the dynamic local torque, effectively eliminating inertial interference caused by the robot's own arm swinging and swaying.

[0046] The temperature drift compensation module uses three platinum resistance temperature measurement points embedded in the sensor base to obtain a discrete temperature field sequence (sampling rate 50Hz). It extracts the average temperature, temperature difference, and heating rate to form an extended temperature feature vector. Through a temperature drift torque coupling model (six-input-nine-output fully connected network) pre-calibrated in a water tank from -5℃ to 40℃, it infers the current six-dimensional force drift estimate. It decouples the drift value from the dynamic local torque to obtain the contact torque magnitude. During the silent period when the robot is stationary and the magnitude is less than 0.3 N·m, the residual zero-point offset is extracted using exponential moving average. If the offset is stable and non-zero for five consecutive silent periods, the global zero-point offset is updated. At the same time, when the temperature difference between adjacent sampling points exceeds 1℃ / s, a temperature shock freezing mechanism is triggered, thereby obtaining a stable local torque that is not affected by seawater temperature fluctuations and zero-point drift.

[0047] The decoupling and transformation module finally combines the stable local torques of each joint into local force spinors according to Fx, Fy, Fz, Mx, My, and Mz. After being projected onto the robot's torso center of mass coordinate system through the adjoint transformation matrix, the global force spinor is obtained. Based on the 1ms sampling point, the module determines the multi-point contact configuration between the wrist and the flange surface in real time, such as "point contact", "slippage" or "pressurization". Based on this, force vector direction semantics such as "alignment error of 0.2mm requires fine adjustment" and "pre-tightening force has reached 85% of the target" and torque transient features such as "bolt slippage signs" are generated. Combined with the current task stage of "underwater flange alignment-pre-tightening-final tightening", the task action semantics are output. Finally, the above multi-level semantics are jointly labeled as global force-sensing voice prompts and digital twin visualization guidance, which successfully guides the underwater remotely operated robot to complete the flange bolt tightening with a torque accuracy of ±0.2N·m, effectively ensuring the safety and reliability of deep-sea oil and gas pipeline connections.

[0048] An electronic device, comprising: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement any of the methods described in the above scheme.

[0049] A storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the six-dimensional force perception space decoupling method based on a humanoid robot as described above.

[0050] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A six-dimensional force perception space decoupling method based on a humanoid robot, characterized in that, include: By synchronously collecting and cleaning the robot's multidimensional strain signals through multidimensional data channels, effective torque data in different dimensions are obtained. By nonlinearly coupling and extending the effective torque data, a static six-dimensional force space is established, and the static local torque is calculated. Based on joint kinematics and the robot's joint motion trajectory, a six-dimensional inertial vector is constructed, and the static local torque is compensated to obtain the dynamic local torque; The dynamic local torque is decoupled from the discrete temperature field sequence and the zero-point spatial drift is suppressed to obtain the stable local torque; The stable local torque is projected and transformed to the robot's center of mass coordinate system, and the global force perception semantics are output.

2. The six-dimensional force perception space decoupling method based on a humanoid robot according to claim 1, characterized in that, The process involves synchronously acquiring and cleaning the robot's multidimensional strain signals through multidimensional data channels to obtain effective torque data in different dimensions, including: Time alignment is performed on the data transmission channel, and a multi-dimensional data channel is constructed by combining data sampling points; Based on the preset data acquisition frequency and the multi-dimensional data channel, the joint strain state of the robot is monitored and the original strain signal is acquired. The original strain signal is filtered and denoised to remove artifacts, resulting in a noise-free strain signal. Peak detection is performed on the noiseless strain signal, the differential amplitude is calculated, and it is compared with a preset standard threshold to determine the peak point and the effective point. If the differential amplitude is greater than the standard threshold, the current data sampling point is determined to be a peak point, and the peak point is filled by linear interpolation of adjacent valid points. If the amplitude of the effective point is greater than the preset full-scale value, the current effective point is determined to be overloaded and marked as an invalid impact, thus obtaining an effective strain signal; Temperature monitoring is performed at data sampling points to collect temperature data and form a local temperature gradient sequence; Based on a lightweight fully connected network, temperature drift inference is performed on the local temperature gradient sequence and running timestamps to calculate the zero-point drift. The effective strain signal is zero-point corrected and analog-to-digital converted based on the zero-point drift to obtain the steady-state strain voltage value. The steady-state strain voltage value is extended and decoupled to obtain effective torque data in different dimensions.

3. The six-dimensional force perception space decoupling method based on a humanoid robot according to claim 1, characterized in that, The nonlinear coupling and extension of the effective torque data establish a static six-dimensional force space, and the calculation of static local torques includes: The effective torque data is normalized based on the full-scale value to obtain the standard torque data; The standard torque data is extended between dimensions to construct an extended feature vector containing linear terms, square terms, and cross-product terms; Based on the nonlinear decoupling matrix, the extended eigenvectors are mapped and reasoned to generate a static six-dimensional force space; Based on a preset local coordinate system reference, the static six-dimensional force space is vector-combined to calculate the static local torque.

4. The six-dimensional force perception space decoupling method based on a humanoid robot according to claim 1, characterized in that, The process of constructing a six-dimensional inertial vector based on joint kinematics and the robot's joint motion trajectory, and compensating for the static local torque to obtain the dynamic local torque, includes: The motion trajectory of the robot's joints is calculated based on joint kinematics to obtain a set of motion parameters; The acceleration of the centroid of the sampling point is calculated based on the centroid position vector of the sampling point and the linear acceleration, angular velocity and angular acceleration in the set of motion parameters. Based on the centroid acceleration at the sampling point and the local gravitational components in the set of motion parameters, the inertial force and inertial torque are calculated, and a six-dimensional inertial vector is synthesized. The original dynamic torque is obtained by performing compensation calculations on the static local torque and the six-dimensional inertia vector. The original dynamic torque is filtered for noise reduction and peak identification to screen out dynamic local torques.

5. The six-dimensional force perception space decoupling method based on a humanoid robot according to claim 1, characterized in that, The step of decoupling the dynamic local torque based on the discrete temperature field sequence and suppressing zero-point spatial drift to obtain a stable local torque includes: The data sampling points are synchronously measured based on the global sampling timestamp to generate a discrete temperature field sequence and extract the extended temperature feature vector. Based on the temperature drift torque coupling model, the extended temperature feature vector is inferred to calculate the six-dimensional force drift estimate; Based on the six-dimensional force drift estimate, the dynamic local torque is decoupled and compensated to obtain the contact torque magnitude. If the robot joint is stationary and the contact torque modulus is less than the preset quiescent threshold, the robot motion state is determined to be a quiescent period without external force, and the current contact torque modulus is subjected to an exponential moving average to extract the residual zero-point offset value. If the residual zero-point offset value remains stable and non-zero for N periods of no external force, it is added to the drift compensation amount as a new zero-point correction term to update the global zero-point offset. If the temperature difference between data sampling points in adjacent sampling periods is greater than the preset temperature threshold, the robot's motion state is determined to be in a temperature shock period, and the compensation weight of the current contact torque modulus is temporarily frozen until the temperature difference converges. After stabilizing the contact torque modulus by both temperature and zero point, a stable local torque is obtained.

6. The six-dimensional force perception space decoupling method based on a humanoid robot according to claim 1, characterized in that, The projection and transformation of the stable local torque to the robot's center-of-mass coordinate system, and the output of global force perception semantics, include: Based on the data dimensions, stable local torques are combined to construct local force spinors in each dimension; The local force spinor is projected onto the robot's center of mass coordinate system using the adjoint transformation matrix to obtain the global force spinor; Based on the data sampling points and the global force rotation, the robot's contact state is determined, and multi-point contact configurations are identified. Based on the multi-point contact configuration, semantic mapping is performed on the stable local torque to extract the semantics of the force vector direction and the transient features of the torque. Based on the current action primitive, the transient characteristics of the torque are associated with the task stage to generate task action semantics; The task action semantics and the force vector direction semantics are labeled to obtain the global force perception semantics.

7. A six-dimensional force perception spatial decoupling system based on a humanoid robot, used to implement the six-dimensional force perception spatial decoupling method based on a humanoid robot as described in any one of claims 1 to 6, characterized in that, include: The data acquisition and cleaning module is used to synchronously acquire and clean the robot's multidimensional strain signals according to the multidimensional data channel, and obtain effective torque data in different dimensions. The static coupling module is used for nonlinear coupling and expansion of the effective torque data, establishing a static six-dimensional force space, and calculating static local torques. The dynamic compensation module is used to compensate for the static local torque based on the joint kinematics and the robot joint motion trajectory to obtain the dynamic local torque; The temperature drift compensation module is used to decouple the dynamic local torque according to the discrete temperature field sequence and suppress zero-point spatial drift to obtain a stable local torque. The decoupling transformation module is used to project and transform the stable local torque to the robot's center of mass coordinate system and output global force perception semantics.

8. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-6.

9. A storage medium storing at least one instruction, at least one program, a code set, or an instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the six-dimensional force perception spatial decoupling method based on a humanoid robot as described in any one of claims 1 to 6.

Images