Complex motion control algorithm and device for double-wheel foot robot

Abstract

The application relates to the technical field of motion control of a double-wheel foot robot, in particular to a complex motion control algorithm and device for a double-wheel foot robot, which comprises a plantar six-dimensional force / torque sensor array, an inertial measurement unit, a joint encoder and a main control unit; the main control unit identifies a support phase and terrain stiffness through a contact state discrimination model, fuses multi-source sensing data to estimate high-precision centroid states through a nonlinear state observer, and generates joint torque instructions by adopting a layered impedance controller. The application can realize high-precision attitude estimation and dynamic balance adjustment in an unstructured terrain, and significantly improves the motion adaptability and anti-disturbance capability of the robot.

Description

Technical Field

[0001] This invention belongs to the field of robot control technology, specifically a complex motion control algorithm and device for a bipedal robot. Background Technology

[0002] Existing motion control methods for bipedal robots mostly employ control strategies based on simplified dynamic models and fixed gait planning. In flat ground environments, they achieve basic walking functions through preset joint trajectories and balance feedback mechanisms, and combine gyroscope and encoder data to adjust posture in order to maintain the stability of the robot during movement.

[0003] However, since such methods usually do not fully consider the abrupt changes in contact force and the coupling effect of multiple degrees of freedom in complex terrain, they are difficult to accurately characterize the foot-ground interaction characteristics when facing unstructured environments such as slopes, steps or soft ground. This makes them susceptible to external disturbances during dynamic adjustment, and may lead to the accumulation of posture estimation bias in high dynamic or large-scale motion scenarios. Ultimately, they fail to effectively guarantee the overall coordination and anti-disturbance ability under complex motion, thus limiting the adaptability of bipedal robots in real complex scenarios. Summary of the Invention

[0004] To this end, embodiments of the present invention provide a complex motion control algorithm and device for biwheeled legs robots. By constructing a real-time motion planning framework based on contact force perception and multi-degree-of-freedom collaborative optimization, and combining a nonlinear state observer and a hierarchical impedance controller, high-precision attitude estimation and dynamic balance adjustment are achieved in unstructured terrain, thereby improving the motion adaptability and anti-disturbance capability of biwheeled legs robots in complex environments.

[0005] In a first aspect, the present invention provides a complex motion control algorithm for a bipedal robot.

[0006] This invention is achieved through the following technical solution: A complex motion control algorithm for a biwheeled legged robot, the algorithm comprising: A six-dimensional force / torque sensor array is installed on the sole of each foot of the bipedal robot. The six-dimensional force / torque sensor array is used to collect foot-ground contact force and torque signals in real time, and the contact force and torque signals are transmitted to the main control unit. The main control unit constructs a contact state discrimination model based on the contact force and torque signals, identifies the current support phase, slip state and terrain stiffness parameters, and inputs the identification results to the nonlinear state observer; The nonlinear state observer integrates the attitude angular velocity and acceleration data output by the inertial measurement unit, the position and velocity information fed back by the joint encoder, and the output of the contact state discrimination model. It uses an extended Kalman filter structure to estimate the position, velocity, and tilt angle of the robot's whole body center of mass online, generating a high-precision state vector. The main control unit calculates the target torque command for each joint based on the high-precision state vector and the preset motion target through a hierarchical impedance controller. The upper controller takes the whole body center of mass trajectory as input and uses the model predictive control method to solve the reference joint trajectory that satisfies the dynamic constraints. The lower controller generates the joint torque command based on the reference joint trajectory and the actual joint state using an adaptive impedance law. The main control unit sends the joint torque command to each joint servo driver to drive the bipedal robot to perform complex motion actions, and continuously receives feedback signals from the six-dimensional force / torque sensor array during the motion to update the contact state discrimination model and the state estimate of the nonlinear state observer.

[0007] In a preferred embodiment of the present invention, the six-dimensional force / torque sensor array is further configured such that a six-dimensional force / torque sensor is embedded in each of the three regions of the foot: the forefoot, heel, and arch. Each sensor is fixed to a preset mounting hole in the metal frame of the foot with bolts, and the sensor output is connected to the analog signal acquisition interface of the main control unit through a shielded cable.

[0008] In a preferred embodiment of the present invention, the construction steps of the contact state discrimination model may further include: The collected contact force and torque signals are low-pass filtered to extract the vertical force component and the horizontal friction force component. Set a vertical force threshold and friction cone boundary conditions. When the vertical force is greater than the threshold and the friction force is located within the friction cone, it is determined to be a stable support state. When the vertical force is less than the threshold or the friction force exceeds the friction cone boundary, it is determined to be a slip or ground-lift state. Based on the state determination results within multiple consecutive sampling periods, a sliding window majority voting mechanism is used to determine the current support phase, and the local stiffness of the terrain is estimated by combining the vertical force change rate.

[0009] In a preferred embodiment of the present invention, the nonlinear state observer may be further configured as follows: The observer gain matrix is ​​adjusted online using the Lyapunov stability condition to ensure that the state estimation remains convergent even under abrupt changes in contact force.

[0010] In a preferred embodiment of the present invention, the optimization problem of the upper-level model predictive control in the hierarchical impedance controller can be further configured as follows: Minimize the weighted sum of centroid trajectory tracking error and control energy consumption in the prediction time domain, with constraints including joint angle limits, joint velocity limits, and ground reaction force feasibility cone; The optimization variable is the joint reference trajectory for several future steps, which is solved online iteratively using a sequential quadratic programming solver.

[0011] This invention provides a complex motion control device for a bipedal robot.

[0012] This invention is achieved through the following technical solution: A complex motion control device for a bipedal robot, the device comprising: A six-dimensional force / torque sensor array is embedded in the forefoot, heel, and arch areas of the left and right feet of the dual-wheeled legged robot to collect foot-ground contact force and torque signals in real time. An inertial measurement unit (IMU) is fixedly installed near the center of gravity of the robot's torso and is used to output the three-axis angular velocity and linear acceleration data of the robot body. Joint encoders, integrated inside each joint motor, are used to provide feedback on joint angle and angular velocity information; The main control unit communicates with the servo drivers of each joint via the CAN bus, receives signals from the six-dimensional force / torque sensor array via the analog input interface, reads data from the inertial measurement unit via the SPI interface, and obtains the joint status via the encoder interface. The main control unit integrates a contact state discrimination module, a nonlinear state observation module, and a layered impedance control module. The contact state discrimination module receives signals from a six-dimensional force / torque sensor array and outputs support phase and terrain stiffness parameters. The nonlinear state observation module fuses multi-source sensor data to generate a high-precision state vector. The layered impedance control module generates joint torque commands based on the state vector and the moving target.

[0013] Thirdly, the present invention provides a computer device.

[0014] This invention is achieved through the following technical solution: A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the aforementioned complex motion control algorithm for any bipedal robot.

[0015] Fourthly, the present invention provides a computer-readable storage medium.

[0016] This invention is achieved through the following technical solution: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned complex motion control algorithm for any bipedal robot.

[0017] In summary, compared with the prior art, the beneficial effects of the technical solution provided by the embodiments of the present invention include at least the following: by embedding a six-dimensional force / torque sensor array in three key areas of the sole, complete mechanical information of foot-ground interaction is directly obtained, avoiding attitude estimation drift caused by relying solely on inertial data; the contact state discrimination model performs multi-cycle sliding window judgment based on vertical force threshold and friction cone boundary conditions, effectively distinguishing stable support, slippage and off-ground states, providing accurate contact event trigger signals for the state observer; the nonlinear state observer integrates contact force, inertial data and joint encoder information, adopts an extended Kalman filter structure and adjusts the observation gain online, significantly improving the accuracy of centroid state estimation in unstructured terrain; the hierarchical impedance controller decouples trajectory planning and force control, the upper-level model predictive control considers dynamic constraints to generate feasible reference trajectories, and the lower-level adaptive impedance law introduces nonlinear compensation terms for contact force, enabling the joint response to both track the trajectory and adapt to external disturbances, thereby maintaining dynamic balance and completing highly coordinated movements in complex environments such as slopes, steps or soft ground. Attached Figure Description

[0018] Figure 1 This is a system structure block diagram of the complex motion control device for the bipedal robot of the present invention, showing the signal connection relationship between each functional module. Detailed Implementation

[0019] Specific embodiments of the complex motion control algorithm and device for the bipedal robot of the present invention are combined with Figure 1 The system architecture block diagram shown is explained in detail. For example... Figure 1 As shown, the complex motion control device for the biwheeled legged robot in this embodiment includes a six-dimensional force / torque sensor array, an inertial measurement unit, a joint encoder, a main control unit, a contact state discrimination module, a nonlinear state observation module, a layered impedance control module, and a joint servo driver. The components are connected through specific physical interfaces and communication protocols, and work collaboratively according to a preset data flow path to achieve real-time control of the biwheeled legged robot's complex motion in unstructured terrain.

[0020] A six-dimensional force / torque sensor array is embedded in three areas of the left and right feet of the bipedal robot: the forefoot, heel, and arch. One six-dimensional force / torque sensor is fixed to each area, for a total of six sensors. Each sensor is bolted to a pre-drilled mounting hole in the foot's metal frame, ensuring a rigid connection and preventing loosening due to vibration or impact. The sensor outputs are led out via shielded cables and connected to the analog signal acquisition interface of the main control unit for transmitting real-time foot-ground contact force and torque signals. This signal includes force components (Fx, Fy, Fz) in three directions and torque components (Mx, My, Mz) in three directions, with a sampling frequency of at least 1 kHz to meet dynamic response requirements.

[0021] The inertial measurement unit (IMU) is a MEMS device integrating a three-axis gyroscope and a three-axis accelerometer, encapsulated in a waterproof and shockproof housing. It is fixedly mounted near the robot's center of gravity to minimize the impact of rotational motion on its output. The IMU connects to the main control unit via an SPI bus, periodically sending the robot's three-axis angular velocity and linear acceleration data to the main control unit, with a typical update frequency of 500 Hz. This data serves as one of the state inputs for the nonlinear state observation module, participating in the fusion estimation of the center of mass motion state.

[0022] The joint encoders are integrated inside each joint servo motor, employing high-resolution absolute encoders with a resolution of 17 bits or higher, accurately feeding back the actual angle and angular velocity information of each joint. All joint encoders are connected to the main control unit via a dedicated encoder interface. The main control unit obtains the position and velocity status of each joint by reading the encoder pulse signals, updating at a frequency synchronized with the control cycle, typically 1 kHz. This information is used not only for error calculation of the lower-level impedance controller but also as input variables for the dynamic model of the nonlinear state observation module.

[0023] The main control unit is the core processing unit, employing a multi-core ARM architecture embedded processor with real-time operating system support and an operating frequency of no less than 1 GHz. The main control unit has multiple hardware interfaces: a CAN bus interface for communication with each joint servo driver; an analog input interface for receiving raw signals from a six-dimensional force / torque sensor array; an SPI interface for reading data from the inertial measurement unit; and an encoder interface for acquiring feedback signals from the joint encoder. Internally, the main control unit integrates three software functional modules: a contact state discrimination module, a nonlinear state observation module, and a hierarchical impedance control module. Each module executes sequentially according to the control cycle, forming a closed-loop control process.

[0024] The contact state discrimination module receives signals from a six-dimensional force / torque sensor array via the analog input interface of the main control unit. First, the raw signal undergoes low-pass filtering with a cutoff frequency set to 30 Hz to remove high-frequency noise interference. Then, the vertical force component Fz and the horizontal friction component √(Fx² + Fy²) are extracted. A vertical force threshold of 5 N is set. When Fz is greater than this threshold and the friction force is less than μ·Fz (where μ is a preset friction coefficient of 0.6), the foot is determined to be in a stable support state. If Fz is less than 5 N, it is determined to be in an off-ground state. If Fz is greater than 5 N but the friction force exceeds the friction cone boundary, it is determined to be in a slip state. The above judgment results are then used for majority voting within a sliding time window (10 sampling periods, or 10 ms) to ultimately determine the current support phase. Simultaneously, the local terrain stiffness parameter k_terrain = ΔFz / Δδ is estimated based on the rate of change of Fz ΔFz / Δt from multiple consecutive sampling points, where Δδ is the foot compression displacement, approximated by the integrated vertical velocity.

[0025] The nonlinear state observation module receives support phase and terrain stiffness parameters from the contact state discrimination module, attitude angular velocity and acceleration data from the inertial measurement unit, and position and velocity information from the joint encoder. The state vector constructed by this module is defined as x = [p_c, v_c, θ, ω]^T, where p_c represents the two-dimensional position (x, z) of the center of mass in the world coordinate system, v_c is its corresponding velocity, θ is the body pitch angle, and ω is the pitch angular velocity. The system dynamic equations are derived based on Lagrange mechanics, considering the gravity term g, the Coriolis term C(q, \dot{q})\dot{q}, and the external contact force term J^T f_ext, where J is the Jacobian matrix and f_ext is the contact force vector provided by the six-dimensional force / torque sensor array. The nonlinear state observation module employs an extended Kalman filter (EKF) structure, using a dynamic model to advance state estimation during the prediction phase and correcting the estimated value based on actual measurement residuals during the update phase. The observer gain matrix K_k is adjusted online using the Lyapunov stability condition. Specifically, the Lyapunov function V = e^TP^{-1}e is constructed, where e is the estimation error and P is the covariance matrix. When a sudden change in contact force is detected (such as crossing a step or stepping into soft ground), the process noise covariance Q is increased, thereby improving the observer's response sensitivity to sudden disturbances and ensuring the convergence of state estimation.

[0026] The hierarchical impedance control module consists of an upper-level model predictive controller and a lower-level adaptive impedance controller. The upper-level controller receives the high-precision state vector x output by the nonlinear state observation module and the user-defined motion target (such as the desired centroid trajectory p_d(t)). Within each control cycle, the upper-level controller solves the following optimization problem within a finite prediction time domain of N=20 steps (5 ms per step, 100 ms in total): minimizing the objective function J = Σ_{k=1}^{N} (||p_c(k) - p_d(k)||²_Q + ||u(k)||²_R), where Q and R are the state error weight matrix and the control input weight matrix, respectively, and u(k) is the joint reference trajectory increment. The constraints include joint angle limits q_min ≤ q(k) ≤ q_max, joint velocity limits |\dot{q}(k)| ≤ \dot{q}_max, and a ground reaction force feasibility cone constraint f_ext ∈ FC, where FC is defined by the friction cone and the upper and lower limits of the vertical force. This optimization problem is solved online iteratively using a sequential quadratic programming (SQP) solver, outputting the joint reference trajectory q_ref = [q_ref(1), ..., q_ref(N)] and its derivatives \dot{q}_ref for the next few steps.

[0027] The lower-level adaptive impedance controller receives the current-time reference trajectory q_ref,i and \dot{q}ref,i output by the upper-level controller, as well as the actual joint states q_i and \dot{q}i fed back by the joint encoder. For the i-th joint, its output torque command τ_i is calculated by the following formula: τ_i = K_d(\dot{q}{ref,i} - \dot{q}i) + K_p(q{ref,i} -q_i) + B_i(f{ext,i}). Where K_p and K_d are diagonal proportional-differential gain matrices, whose values ​​are pre-tuned according to the joint type; B_i(·) is a nonlinear compensation term, whose functional form is obtained through offline system identification, specifically a polynomial fitting function B_i(f) = a_0 + a_1 f + a_2 f² + a_3 f³, and the coefficients a_j are obtained by fitting experimental data using the least squares method. This compensation term directly reflects the mapping relationship between contact force and external disturbance, enabling the joint to actively adjust the output torque to adapt to terrain changes while tracking the reference trajectory.

[0028] The main control unit sends the torque commands τ_i generated by the hierarchical impedance control module to the corresponding joint servo drivers via the CAN bus. Upon receiving the torque commands, the joint servo drivers drive the corresponding joint motors to generate the required output torque, thus driving the linkage mechanism to complete the leg movement. During this process, the six-dimensional force / torque sensor array continuously collects new contact force signals and transmits them back to the main control unit via the analog input interface for contact state discrimination and state estimation updates in the next control cycle, forming a complete closed-loop control link.

[0029] In practical applications, such as when a bipedal robot walks on a sloping terrain, the six-dimensional force / torque sensor array detects a significant increase in the force (Fz) in the forefoot region, indicating that the friction force is close to the friction cone boundary. Based on this, the contact state discrimination module determines that the robot is in a stable support position but has a potential risk of slippage. The nonlinear state observation module integrates this information with inertial data to correct the estimated center of mass tilt angle, preventing attitude misjudgment caused by slope inclination. The upper-level MPC of the layered impedance control module adjusts the reference trajectories of the hip and knee joints to keep the center of mass projection within the support polygon; the lower-level impedance controller enhances ankle joint damping through the B_i(f_ext,i) term to suppress high-frequency jitter caused by uneven slope. The entire control cycle is executed cyclically at a frequency of 1 kHz, ensuring that the robot maintains dynamic balance and completes highly coordinated gait movements in complex environments.

[0030] The connections between the aforementioned components are clearly defined: the six-dimensional force / torque sensor array is connected to the analog input interface of the main control unit via a shielded cable; the inertial measurement unit is connected to the main control unit via an SPI bus; the joint encoder is connected to the main control unit via an encoder interface; and the main control unit is connected to each joint servo driver via a CAN bus. The contact state discrimination module, nonlinear state observation module, and hierarchical impedance control module within the main control unit are executed sequentially in series, forming a complete control logic chain. Data transmission between modules is completed within a single control cycle, ensuring real-time performance. All hardware selections and parameter settings are based on existing industrial-grade component specifications, requiring no special customization to implement the aforementioned technical solution.

[0031] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principles of this invention are further supplemented below with a specific application scenario.

[0032] When the bipedal robot traverses a 15 cm high step, the six-dimensional force / torque sensor array first detects that the vertical force component Fz in the forefoot region rapidly increases from 0 N to 8 N the instant the swing leg's toe contacts the upper edge of the step. Simultaneously, the horizontal friction component √(Fx² + Fy²) reaches 4.9 N, approaching the preset friction cone boundary μ·Fz = 0.6 × 8 = 4.8 N. After receiving this signal, the contact state discrimination module processes it through a 30 Hz low-pass filter and determines that the foot is in a critical support state within 10 consecutive sampling periods (10 ms). Combining the Fz rise slope ΔFz / Δt, it estimates that the step contact stiffness k_terrain is significantly higher than that of a flat surface, thus outputting a "high stiffness contact" flag and a support phase switching command to the nonlinear state observation module.

[0033] The nonlinear state observation module, based on the contact event-triggered state update mechanism, temporarily increases the process noise covariance Q to three times its original value to enhance the sensitivity of the extended Kalman filter to sudden external forces. Simultaneously, it uses the ankle joint angle abrupt change information fed back from the joint encoder and the trunk pitch acceleration output by the inertial measurement unit to correct the estimated center-of-gravity height p_c,z, avoiding an underestimation of the center-of-gravity position due to sudden foot lift. During this process, the pitch angle θ in the state vector x = [p_c, v_c, θ, ω]^T is corrected in real time to prevent attitude integral drift caused by the step impact.

[0034] The upper-level model predictive controller of the layered impedance control module receives the updated high-precision state vector x and the preset centroid climbing trajectory p_d(t), and re-solves the optimization problem within the 100 ms prediction time domain. The ground reaction force feasibility cone FC in the constraint conditions is dynamically reduced according to the current contact area (only the forefoot is under force) to ensure that the generated hip and knee joint reference trajectories q_ref do not violate the force balance condition under single-point support; the output q_ref(1) is the target joint angle of the current control cycle. The lower-level adaptive impedance controller calculates the ankle joint torque command accordingly. The Bankle(f_ext) term activates nonlinear compensation due to f_ext,Fz = 8N, and its value is determined by the offline-identified cubic polynomial. Calculations show that actively increasing the ankle joint output torque can resist the overturning torque at the edge of the step, while simultaneously increasing local damping to suppress high-frequency oscillations caused by toe impact.

[0035] The main control unit sends τ_ankle and other joint torque commands to the corresponding joint servo drivers via the CAN bus, driving the ankle joint motor to generate precise torque, ensuring the sole of the foot smoothly contacts the upper surface of the step. Simultaneously, the six-dimensional force / torque sensor array of the supporting leg continuously monitors whether the heel region Fz drops below 5 N. If Fz < 5 N and persists for more than 5 ms, the contact state discrimination module determines that the supporting leg has entered the ground-lift preparation phase and notifies the nonlinear state observation module in advance to switch the contact force term J^T f_ext in the dynamic model to zero, achieving a seamless transition between the support and swing phases. The entire control process is executed cyclically at a control frequency of 1 kHz, ensuring that the robot's center-of-gravity trajectory tracking error is less than 2 cm during step crossing, and that the body pitch angle fluctuation is controlled within ±3°, effectively preventing imbalance and falls caused by sudden changes in contact force.

[0036] All contents not described in detail in the specification are existing technologies known to those skilled in the art, and the model parameters of each electrical appliance are not specifically limited; conventional equipment can be used. Electrical control components not mentioned in this technical solution are not shown in the figures because they are existing technologies, and will not be described here.

[0037] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A complex motion control algorithm for a bipedal robot, characterized in that, The algorithm includes: A six-dimensional force / torque sensor array (1) is installed on the sole of each foot of the two-wheeled leg robot. The six-dimensional force / torque sensor array (1) is used to collect foot-ground contact force and torque signals in real time, and the contact force and torque signals are transmitted to the main control unit (4). The main control unit (4) constructs a contact state discrimination model based on the contact force and torque signals, identifies the current support phase, slip state and terrain stiffness parameters, and inputs the identification results to the nonlinear state observer; The steps for constructing the contact state discrimination model include: The collected contact force and torque signals are low-pass filtered to extract the vertical force component and the horizontal friction force component. Set a vertical force threshold and friction cone boundary conditions. When the vertical force is greater than the threshold and the friction force is located within the friction cone, it is determined to be a stable support state. When the vertical force is less than the threshold or the friction force exceeds the friction cone boundary, it is determined to be a slip or ground-lift state. Based on the state determination results within multiple consecutive sampling periods, a sliding window majority voting mechanism is used to determine the current support phase, and the local stiffness of the terrain is estimated by combining the vertical force change rate. The contact force and torque signals include force components (Fx, Fy, Fz) in three directions and torque components (Mx, My, Mz) in three directions, with a sampling frequency of not less than 1kHz. The nonlinear state observer integrates the attitude angular velocity and acceleration data output by the inertial measurement unit (2), the position and velocity information fed back by the joint encoder (3), and the output of the contact state discrimination model. It uses an extended Kalman filter structure to estimate the position, velocity, and tilt angle of the robot's whole body mass online and generate a high-precision state vector. The main control unit (4) calculates the target torque command of each joint through a hierarchical impedance controller based on the high-precision state vector and the preset motion target. The upper controller takes the whole body center of mass trajectory as input and uses the model predictive control method to solve the reference joint trajectory that satisfies the dynamic constraints. The lower controller generates the joint torque command based on the reference joint trajectory and the actual joint state using an adaptive impedance law. The main control unit (4) sends the joint torque command to each joint servo driver (8) to drive the bipedal robot to perform complex motion actions, and continuously receives feedback signals from the six-dimensional force / torque sensor array (1) during the motion process to update the contact state discrimination model and the state estimation of the nonlinear state observer; The optimization problem of the upper-level model predictive control in the hierarchical impedance controller is: Minimize the weighted sum of centroid trajectory tracking error and control energy consumption in the prediction time domain, with constraints including joint angle limits, joint velocity limits, and ground reaction force feasibility cone; The optimization variable is the joint reference trajectory for several future steps, which is solved online iteratively using a sequential quadratic programming solver.

2. The complex motion control algorithm for a bipedal robot according to claim 1, characterized in that, The six-dimensional force / torque sensor array (1) is installed on the sole of the foot as follows: a six-dimensional force / torque sensor is embedded in the three areas of the forefoot, heel and arch of the foot. Each sensor is fixed in the preset mounting hole of the metal frame of the foot by bolts. The sensor output is connected to the analog signal acquisition interface of the main control unit (4) through a shielded cable.

3. The complex motion control algorithm for a bipedal robot according to claim 1, characterized in that, The nonlinear state observer adopts the following structure: The system's dynamic equations are derived from Lagrange mechanics and include gravity terms, Coriolis terms, and external contact force terms. The observer gain matrix is ​​adjusted online using the Lyapunov stability condition.

4. A complex motion control device for a bipedal robot, characterized in that, The device includes: A six-dimensional force / torque sensor array (1) is embedded in the forefoot, heel and arch area of ​​the left and right feet of the two-wheeled leg robot to collect foot-ground contact force and torque signals in real time. An inertial measurement unit (2) is fixedly installed near the center of gravity of the robot's torso and is used to output the three-axis angular velocity and linear acceleration data of the robot body; The joint encoder (3) is integrated inside each joint motor and is used to feed back joint angle and angular velocity information; The main control unit (4) communicates with each joint servo driver (8) via the CAN bus, receives signals from the six-dimensional force / torque sensor array (1) via the analog input interface, reads data from the inertial measurement unit (2) via the SPI interface, and obtains information from the joint encoder (3) via the encoder interface. The main control unit (4) integrates a contact state discrimination module (5), a nonlinear state observation module (6), and a layered impedance controller. The contact state discrimination module (5) receives signals from the six-dimensional force / torque sensor array (1) and outputs support phase and terrain stiffness parameters. The nonlinear state observation module (6) fuses multi-source sensor data to generate a high-precision state vector. The layered impedance controller generates joint torque commands based on the state vector and the moving target. The main control unit (4) sends the joint torque command to each joint servo driver (8) to drive the bipedal robot to perform complex motion actions, and continuously receives feedback signals from the six-dimensional force / torque sensor array (1) during the motion process to update the contact state discrimination model and the state estimation of the nonlinear state observer; The steps for constructing the contact state discrimination model include: The collected contact force and torque signals are low-pass filtered to extract the vertical force component and the horizontal friction force component. Set a vertical force threshold and friction cone boundary conditions. When the vertical force is greater than the threshold and the friction force is located within the friction cone, it is determined to be a stable support state. When the vertical force is less than the threshold or the friction force exceeds the friction cone boundary, it is determined to be a slip or ground-lift state. Based on the state determination results within multiple consecutive sampling periods, a sliding window majority voting mechanism is used to determine the current support phase, and the local stiffness of the terrain is estimated by combining the vertical force change rate. The contact force and torque signals include force components (Fx, Fy, Fz) in three directions and torque components (Mx, My, Mz) in three directions, with a sampling frequency of not less than 1kHz. The optimization problem of the upper-level model predictive control in the hierarchical impedance controller is: Minimize the weighted sum of centroid trajectory tracking error and control energy consumption in the prediction time domain, with constraints including joint angle limits, joint velocity limits, and ground reaction force feasibility cone; The optimization variable is the joint reference trajectory for several future steps, which is solved online iteratively using a sequential quadratic programming solver.

5. The complex motion control device for a bipedal robot according to claim 4, characterized in that, The six-dimensional force / torque sensor array (1) includes six six-dimensional force / torque sensors, three on each of the left and right soles, located in the forefoot, heel and arch areas respectively. Each sensor is fixed to the preset mounting holes of the metal skeleton on the sole of the foot by bolts. The sensor output ends are connected to the analog signal acquisition interface of the main control unit (4) through shielded cables.

Images