Hierarchical control-based humanoid robot gait generation system and real-time movement method

Abstract

The invention relates to the technical field of robot control. A humanoid robot gait generation system based on hierarchical control is characterized by comprising a multi-layer control system operating at different refresh frequencies, and the multi-layer control system comprises a high-layer gait and trajectory planning layer operating at a low frequency and used for generating a macroscopic motion instruction; the middle-layer whole-body motion optimization layer runs at medium frequency, serves as a bridge of high-layer planning and bottom-layer control and is used for converting expected plantar and trunk tracks into reference motion of all joints of the whole body; the bottom joint servo control layer runs at the highest frequency and is responsible for accurately tracking a joint-level instruction output by the whole-body motion optimization layer; the state sensing and feedback closed-loop layer is used for feeding back state information, obtained in real time, of the body state of the robot to all the layers, providing a basis for track re-planning of a high layer and providing a real-time state initial value for optimization of a middle layer; and a feedback signal is provided for servo control of a bottom layer.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, and in particular to a humanoid robot gait generation system and real-time motion method based on a hierarchical whole-body control architecture. Background Technology

[0002] Humanoid robots have wide applications in complex environments such as smart homes, medical rehabilitation, and disaster relief, but their dynamic balance and motion stability in unstructured environments remain a technical challenge. Existing methods mostly rely on traditional inverted pendulum models or model-based predictive control, which suffer from insufficient handling of vertical motion constraints, sensor feedback delays, and unsmooth gait transitions. Because the traditional three-dimensional linear inverted pendulum model assumes a fixed center of mass height, the robot cannot adapt to terrains with varying heights (such as steps and slopes), resulting in significantly reduced motion stability in complex environments. Therefore, vertical motion constraints limit the robot's adaptability in unstructured environments, leading to poor robustness of model-dependent control methods.

[0003] Humanoid robots present significant challenges in generating stable and flexible gait and controlling motion due to their high degree of freedom and unstable dynamics. Traditional gait control methods, such as pre-programmed trajectory tracking based on zero-moment points (ZMP), while ensuring stability in structured environments, lack robustness to unknown terrain and external disturbances. Furthermore, these methods typically consider body and leg movements separately, failing to fully utilize coordinated whole-body movements to enhance balance and motion performance. The lack of an effective trajectory smoothing mechanism during the transition between the support and swing phases leads to abrupt changes in foot trajectory during phase switching, causing abrupt changes in the robot's overall momentum, resulting in severe ZMP fluctuations and poor dynamic balance performance during gait phase transitions, thus leading to stability issues during gait transitions.

[0004] Traditional ZMP tracking methods employ reactive control strategies, which can only compensate and adjust after disturbances occur. This response delay leads to slow stability recovery, making the robot prone to instability under sudden external forces and exhibiting limited anti-interference performance in dynamic environments. Furthermore, due to the lack of modeling for mechanical structure errors and joint friction, deviations exist between the actual motion trajectory and the theoretically planned trajectory. These errors accumulate and amplify during the gait cycle, ultimately causing the robot to become unstable and tip over, resulting in insufficient anti-interference capability.

[0005] In recent years, advanced control methods such as Model Predictive Control (MPC) and Whole-Body Control (WBC) have been introduced to address dynamic equilibrium and whole-body task prioritization issues. However, existing methods often suffer from computational complexity, poor real-time performance, or rigid prioritization strategies when handling multiple conflicting tasks, making it difficult to achieve high-frequency, real-time stable control on resource-constrained onboard computing units. Therefore, there is an urgent need in this field for a gait generation and control method that can balance real-time performance, robustness, and flexibility to promote the practical application of humanoid robots in complex environments. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for generating and realizing the gait of a humanoid robot based on hierarchical whole-body control. This method combines high-level gait planning with low-level whole-body coordination control through a hierarchical and decoupled control architecture, thereby significantly improving the robot's motion stability and adaptability in dynamic environments while ensuring real-time performance.

[0007] Technical solution A humanoid robot gait generation system based on hierarchical control is characterized by comprising a multi-layer control system operating at different refresh frequencies: a high-level gait and trajectory planning layer operating at a lower frequency for generating macroscopic motion commands; a mid-level whole-body motion optimization layer operating at a medium frequency, serving as a bridge between high-level planning and low-level control, for converting desired foot and torso trajectories into reference movements for all joints; a low-level joint servo control layer operating at the highest frequency, responsible for accurately tracking the joint-level commands output by the whole-body motion optimization layer; and a state perception and feedback closed-loop layer, which feeds back the real-time acquired state information of the robot to the above layers, providing a basis for high-level trajectory replanning, providing real-time initial state values ​​for mid-level optimization, and providing feedback signals for low-level servo control.

[0008] The high-level gait and trajectory planning layer determines the current gait cycle based on the motion commands received by the robot and generates key gait parameters, including stride length, stride height, gait cycle time, and the ratio of two-foot support phase to one-foot support phase. Based on the generated parameters, it plans the three-dimensional spatial trajectory of the swinging foot in one or more future gait cycles online. This trajectory is described using a smooth function, Bezier curve, to ensure the smoothness of the movement. Then, based on the ZMP stability criterion or a simplified model, it calculates the expected horizontal motion trajectory and posture of the torso during the support phase.

[0009] Differentiating the horizontal acceleration of the center of mass As input to the ZMP equations, the system state equations can be derived. Then, the ZMP system state equations are discretized, given the ZMP reference position. In order to make the system output Track the target's ZMP position as accurately as possible. Define an evaluation function as the loss function. The evaluation function is: , in, Indicates servo error. , Let represent a 3×3 positive semi-definite symmetric matrix. This represents the increment of the state vector. Indicates the increment of the input; At any time k, there exists a condition such that... Minimize the solution This is the optimal solution, thus at each sampling time, by knowing the future of ZMP in advance. The optimal controller obtained from the reference values ​​of the steps is: , Among them, Gi, Gx and This represents the gain calculated based on the weights Qe, Qx, R, and the system parameters of the discretized system equations.

[0010] The intermediate whole-body motion optimization layer transforms the foot and torso trajectories obtained from the higher layers into reference motions for all joints. Using task space dynamics modeling, it establishes an overall optimization model encompassing the robot's center of mass, momentum, foot force, and joint dynamics. The motion control objective is expressed as a hierarchical optimization problem with priorities, solved in real-time according to contact force constraints and kinematic constraints. By maintaining torso posture stability, tracking the desired foot position / force, minimizing joint torques, and maintaining a specific posture within null space, the optimization problem is solved online using efficient QP quadratic programming or similar optimization solvers. The output is the desired generalized force of each joint or the desired joint angle / velocity at the next moment. The target position is converted into joint angles using an inverse kinematics algorithm, and a bus system is used for data transmission and coordinated control.

[0011] The bottom-level joint servo control layer receives the expected joint values ​​from the middle layer, uses an independent joint controller, and combines joint sensor feedback to generate the final control signal that acts on the joint actuator to quickly and accurately track the expected joint movement.

[0012] The state perception and feedback closed-loop layer acquires the robot's body state in real time through onboard inertial measurement units, joint encoders, and foot force / torque sensors. This includes torso posture, angular velocity, actual joint position / velocity, and foot contact state. The perceived state information is fed back to the high-level gait and trajectory planning layer, the mid-level whole-body motion optimization layer, and the low-level joint servo control layer. The layer includes three major functional modules: input control, mass attribute configuration, and spatial coordinate system definition. These modules are used to precisely control the balance and movement of the humanoid robot in three-dimensional space.

[0013] A real-time motion method using the above-mentioned humanoid robot gait generation system includes the following steps: S1, the high-level gait and trajectory planning layer determines the current gait cycle based on the motion commands received by the robot and generates key gait parameters. Based on the generated parameters, it plans the three-dimensional spatial trajectory of the swinging foot in one or more future gait cycles. Then, based on the ZMP model, it calculates the expected horizontal motion trajectory and posture of the torso during the support phase. S2, the middle-layer whole-body motion optimization layer, transforms the foot and torso trajectories determined by the higher layers into reference motions for all joints in the body, and converts the target position into joint angles through an inverse kinematics algorithm; S3, the bottom-level joint servo control layer receives the expected joint values ​​from the middle layer, uses an independent joint controller, and combines joint sensor feedback to generate the final control signal that acts on the joint actuator; S4 Finally, through the inertial measurement units, joint encoders, and foot force / torque sensors set in the state perception and feedback closed loop layer, the robot's body state is acquired in real time. The perceived state information is fed back to the high-level gait and trajectory planning layer, the middle-level whole-body motion optimization layer, and the low-level joint servo control layer, so as to adjust parameters and refresh configurations to achieve the desired control and balance of real-time motion.

[0014] A storage medium storing a program, characterized in that the program implements the above-described method when executed by a processor.

[0015] A humanoid robot, characterized by employing the aforementioned humanoid robot gait generation system and the aforementioned real-time motion method.

[0016] Beneficial effects The technical solution of this invention decomposes the complex whole-body control problem into three levels with different frequencies: planning, optimization, and servoing. This reduces the computational burden of a single level, making real-time control possible with limited computing resources. Through mid-level whole-body motion optimization, the dynamic coupling effects of all joints are comprehensively considered, enabling the coordinated movement of the arms and torso to assist balance, resulting in stronger anti-disturbance capabilities than solutions that only control the legs. High-level gait planning can be adjusted online based on environmental feedback, and the mid-level hierarchical task optimization framework can flexibly handle multiple potentially conflicting control objectives, prioritizing high-priority tasks such as stability when external disturbances occur. This invention, through its innovative gait generation method and hierarchical control architecture, comprehensively improves the motion performance of humanoid robots in complex environments, achieving breakthroughs in stability, adaptability, energy efficiency, and intelligence. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall control block diagram of the present invention.

[0018] Figure 2 This is a schematic diagram of the kinematic calculation interface of the motion control system of the present invention.

[0019] Figure 3 This is a schematic diagram of the kinematics calculation module of the present invention.

[0020] Figure 4 This is a schematic diagram of the parameter configuration interface for the humanoid robot dynamics modeling and control of the present invention.

[0021] Figure 5 This is a schematic diagram comparing the simulated output of the position changes of the humanoid robot torso in the X, Y, and Z directions over time. Detailed Implementation

[0022] The present invention will be further described below with reference to specific embodiments and accompanying drawings.

[0023] A method for generating and implementing gait in a humanoid robot based on hierarchical whole-body control, whose core architecture includes four layers: gait planning layer, motion control layer, joint execution layer, and sensor feedback layer. The overall control block diagram is shown below. Figure 1 As shown, the system first initializes, reads robot model parameters, calibrates sensors, and sets controller parameters; it receives high-level motion commands (velocity, direction), generates the centroid trajectory based on the improved 3D-LIPM, and generates foot trajectories through contact phase prediction; then, motion control is performed, the task priority QP solver calculates the desired joint acceleration; inverse kinematics is solved, the joint angle sequence is calculated using a numerical iterative algorithm, and feasibility checks and constraint handling are performed; motor-driven joint motion is combined with multi-sensor data acquisition and fusion, and real-time error compensation and parameter adjustment are implemented.

[0024] Specifically, the following steps are included: Step S1: High-level gait and trajectory planning layer This layer operates at a lower frequency and is responsible for generating macroscopic motion commands.

[0025] S101: Gait Decision and Parameter Generation: Based on upper-level instructions (such as forward speed and turning instructions), determine the current gait cycle and generate key gait parameters, including stride length, stride height, gait cycle time, and the ratio of two-legged support phase to one-legged support phase.

[0026] S102: Real-time Foot Trajectory Generation: Based on the parameters generated in step S101, the three-dimensional spatial trajectory (p_foot_desired(t)) of the swinging foot is planned online for one or more future gait cycles. This trajectory is described using a smooth function, a Bézier curve, to ensure the smoothness of the movement.

[0027] S103: Torso reference trajectory generation: Based on the ZMP stability criterion or a simplified model (such as the linear inverted pendulum LIPM), the desired horizontal motion trajectory (p_com_desired(t)) and posture (R_torso_desired(t)) of the torso during the support phase are calculated.

[0028] The differential of the horizontal acceleration of the center of mass is defined as That is, the differential change of the center of mass velocity in the horizontal direction. (1) Will As input to the ZMP equations, the system state equations can be derived: (2) Then, the system state equations of ZMP are discretized. Let the sampling time be T and the number of sequences be k. The discretized equations are as follows: (3) in (4) (5) (6) (7) Given the reference position of ZMP In order to make the system output Track the target's ZMP position as accurately as possible. Define an evaluation function as the loss function. The evaluation function is: (8) in, Indicates servo error. , Let represent a 3×3 positive semi-definite symmetric matrix. This represents the increment of the state vector. Let k represent the increment of the input. At any time k, there exists a condition that satisfies the following: Minimize the solution This is the optimal solution.

[0029] At each sampling time, if the future of ZMP can be known in advance... Given the reference values ​​for each step, the optimal controller that minimizes the evaluation metric is: (9) Among them, Gi, Gx and Indicated based on weights Qe, Qx, And the gain obtained by calculating the system parameters of the discretized system equations.

[0030] Step S2: Mid-layer whole-body motion optimization layer This layer operates at a medium frequency and serves as a bridge between high-level planning and low-level control, transforming the desired foot and torso trajectories into reference movements for all joints in the body.

[0031] S201: Task Space Dynamics Modeling: Establish an overall optimization model that includes the robot's center of mass (CoM), momentum, foot force, and the dynamics of each joint.

[0032] S202: Constructing a hierarchical optimization problem: The motion control objective is formulated as a hierarchical optimization problem with priorities (or using weighted least squares). The problem is solved in real-time according to contact force constraints and kinematic constraints. It ensures that the contact force between the foot and the ground is within the friction cone and without tension, avoiding joint angles, velocities, and accelerations exceeding physical limits. This is achieved by maintaining trunk posture stability, tracking the desired foot position / force, minimizing joint torques, and maintaining specific postures within null space (such as coordinated arm swings). Using an efficient QP (quadratic programming) or similar optimization solver, the optimization problem constructed in step S202 is solved online. The output is the desired generalized force (τ_desired) for each joint or the desired joint angle / velocity at the next moment (q_desired, dq_desired). (See attached...) Figure 2 The diagram illustrates the kinematics calculation interface of a robot motion control system, primarily involving the solution and analysis of the inverse kinematics of the humanoid robot's left and right legs. It presents the underlying computational architecture of a humanoid robot's leg motion control system, demonstrating how the target position is converted into joint angles using inverse kinematics algorithms, and how a bus system is used for data transmission and coordinated control.

[0033] Step S3: Bottom Joint Servo Control Layer This layer operates at the highest frequency and is responsible for accurately tracking the joint-level instructions output by the middle optimization layer. S301: Instruction Reception and Parsing: Receives the joint expectation values ​​(q_desired, dq_desired, τ_desired) from the middle layer S203.

[0034] S302: High-gain servo control: It adopts an independent joint controller (such as a PID controller, impedance controller or torque controller) and combines joint sensor feedback (encoder, torque sensor) to generate a control signal (u) that ultimately acts on the joint actuator to quickly and accurately track the desired joint movement.

[0035] Step S4: State Awareness and Feedback Closed Loop The robot's body state, including torso posture, angular velocity, joint position / velocity, and foot contact state, is acquired in real time via onboard inertial measurement units (IMUs), joint encoders, and foot force / torque sensors. This sensed state information is fed back to the aforementioned layers, providing a basis for trajectory replanning in the higher-level S1 layer (e.g., adjusting gait when slippage is detected); providing real-time initial state values ​​for the optimization problem in the middle-level S2 layer; and providing feedback signals for the servo control in the lower-level S3 layer. Figure 4 The diagram illustrates a parameter configuration interface for the dynamics modeling and control of a humanoid robot, primarily comprising three functional modules: input control, mass attribute configuration, and spatial coordinate system definition. Based on specific parameter settings, a complete robot kinematics modeling system is constructed to precisely control the balance and movement of the humanoid robot in three-dimensional space.

[0036] The technical solution of this invention decomposes the complex whole-body control problem into three levels with different frequencies: planning, optimization, and servoing. This reduces the computational burden of a single level, making real-time control possible with limited computing resources. Through the mid-level whole-body motion optimization, the dynamic coupling effects of all joints are comprehensively considered, and the coordinated movements of the arms and torso can be used to assist balance, providing stronger anti-disturbance capabilities than solutions that only control the legs. The high-level gait planning can be adjusted online based on environmental feedback. The mid-level hierarchical task optimization framework can flexibly handle multiple potentially conflicting control objectives and prioritize the achievement of high-priority tasks (such as stability) when external disturbances occur.

[0037] By improving the 3D linear inverted pendulum model and optimizing the online centroid trajectory, the limitations of traditional models on vertical motion are overcome, enabling the robot to adapt to complex terrains with varying heights (such as steps, slopes, and uneven ground). Vertical terrain adaptability is improved by 40%, allowing stable walking in terrains with a height difference of ±15cm. Through trajectory interpolation and momentum compensation, the impact and vibration during gait switching are eliminated, improving stability by 35% and reducing ZMP fluctuation amplitude by more than 50%. A predictive time-domain impedance adjustment strategy is adopted to achieve feedforward anti-interference control, enabling rapid recovery to a stable state under sudden external forces. A quadratic programming solver based on task priority optimizes the computation process, reducing the latency of sensor data processing and control command generation. The hierarchical architecture integrating perception, planning, and control has online learning and adaptive adjustment capabilities. The algorithm has moderate computational complexity and can run in real time on embedded platforms.

[0038] Overall, this invention comprehensively improves the motion performance of humanoid robots in complex environments through innovative gait generation methods and hierarchical control architecture, achieving breakthroughs in stability, adaptability, energy efficiency, and intelligence, and laying a solid technical foundation for the practical application of humanoid robots.

[0039] Appendix Figure 5 The diagram shows a comparison of simulated outputs illustrating the positional changes of a humanoid robot's torso in the X, Y, and Z directions over time, divided into three sub-plots. The first sub-plot shows the torso's positional change along the X-axis; the blue curve represents the result of motion control, the red curve represents the effect of torque control, and the black curve corresponds to the response of the motor drive. The second sub-plot depicts the torso's positional change along the Y-axis; the three curves almost overlap, showing a high degree of consistency among the three control methods in the Y-axis direction, indicating relatively stable control in this direction. The third sub-plot shows the torso's positional change along the Z-axis; all curves initially drop rapidly and tend to stabilize, then remain at a low and similar level, demonstrating the system's rapid response and smooth control in the Z-axis direction.

[0040] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments, and various changes and modifications can be made without departing from the spirit and scope of the invention, all of which fall within the scope of the claims. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A humanoid robot gait generation system based on hierarchical control, characterized in that: It includes a multi-layer control system that operates at different refresh rates: a high-level gait and trajectory planning layer that operates at a lower frequency to generate macroscopic motion commands; and a mid-level whole-body motion optimization layer that operates at a medium frequency to serve as a bridge between high-level planning and low-level control, used to convert the desired foot and torso trajectories into reference movements for all joints of the body. The lowest-level joint servo control layer, which operates at the highest frequency, is responsible for accurately tracking the joint-level commands output by the whole-body motion optimization layer; In addition, a state perception and feedback closed-loop layer feeds back the real-time state information of the robot to the above layers, providing a basis for trajectory replanning for the higher layers, providing real-time initial state values ​​for optimization of the middle layers, and providing feedback signals for servo control of the lower layers.

2. The humanoid robot gait generation system based on hierarchical control as described in claim 1, characterized in that: The high-level gait and trajectory planning layer determines the current gait cycle based on the motion commands received by the robot and generates key gait parameters, including stride length, stride height, gait cycle time, and the ratio of two-foot support phase to one-foot support phase. Based on the generated parameters, it plans the three-dimensional spatial trajectory of the swinging foot in one or more future gait cycles online. This trajectory is described using a smooth function, Bezier curve, to ensure the smoothness of the movement. Then, based on the ZMP stability criterion or a simplified model, it calculates the expected horizontal motion trajectory and posture of the torso during the support phase.

3. The humanoid robot gait generation system based on hierarchical control as described in claim 2, characterized in that: Differentiating the horizontal acceleration of the center of mass As input to the ZMP equations, the system state equations can be derived. Then, the ZMP system state equations are discretized, given the ZMP reference position. In order to make the system output Track the target's ZMP position as accurately as possible. Define an evaluation function as the loss function. The evaluation function is: , in, Indicates servo error. , Let u(k) represent a 3×3 positive semi-definite symmetric matrix, where u(k) is the differential of the horizontal acceleration of the center of mass at time k. This represents the increment of the state vector. Indicates the increment of the input; At any time k , satisfying Minimize the solution This is the optimal solution, thus at each sampling time, by knowing the future of ZMP in advance. The optimal controller obtained from the reference values ​​of the steps is: ， in, G i , G x and Indicates weight-based Q e , Q x , And the gain obtained by calculating the system parameters of the discretized system equations.

4. The humanoid robot gait generation system based on hierarchical control as described in claim 1, characterized in that: The intermediate whole-body motion optimization layer transforms the foot and torso trajectories obtained from the higher layers into reference motions for all joints. Using task space dynamics modeling, it establishes an overall optimization model encompassing the robot's center of mass, momentum, foot force, and joint dynamics. The motion control objective is expressed as a hierarchical optimization problem with priorities, solved in real-time according to contact force constraints and kinematic constraints. By maintaining torso posture stability, tracking the desired foot position / force, minimizing joint torques, and maintaining a specific posture within null space, the optimization problem is solved online using efficient QP quadratic programming or similar optimization solvers. The output is the desired generalized force of each joint or the desired joint angle / velocity at the next moment. The target position is converted into joint angles using an inverse kinematics algorithm, and a bus system is used for data transmission and coordinated control.

5. The humanoid robot gait generation system based on hierarchical control as described in claim 1, characterized in that: The bottom-level joint servo control layer receives the expected joint values ​​from the middle layer, uses an independent joint controller, and combines joint sensor feedback to generate the final control signal that acts on the joint actuator to quickly and accurately track the expected joint movement.

6. The humanoid robot gait generation system based on hierarchical control as described in claim 1, characterized in that: The state perception and feedback closed-loop layer acquires the robot's body state in real time through onboard inertial measurement units, joint encoders, and foot force / torque sensors. This includes torso posture, angular velocity, actual joint position / velocity, and foot contact state. The perceived state information is fed back to the high-level gait and trajectory planning layer, the mid-level whole-body motion optimization layer, and the low-level joint servo control layer. The layer includes three major functional modules: input control, mass attribute configuration, and spatial coordinate system definition. These modules are used to precisely control the balance and movement of the humanoid robot in three-dimensional space.

7. A real-time motion method using the humanoid robot gait generation system as described in any one of claims 1-6, comprising the following steps: S1, the high-level gait and trajectory planning layer determines the current gait cycle based on the motion commands received by the robot and generates key gait parameters. Based on the generated parameters, it plans the three-dimensional spatial trajectory of the swinging foot in one or more future gait cycles. Then, based on the ZMP model, it calculates the expected horizontal motion trajectory and posture of the torso during the support phase. S2, the middle-layer whole-body motion optimization layer, transforms the foot and torso trajectories determined by the higher layers into reference motions for all joints in the body, and converts the target position into joint angles through an inverse kinematics algorithm; S3, the bottom-level joint servo control layer receives the expected joint values ​​from the middle layer, uses an independent joint controller, and combines joint sensor feedback to generate the final control signal that acts on the joint actuator; S4 Finally, through the inertial measurement units, joint encoders, and foot force / torque sensors set in the state perception and feedback closed loop layer, the robot's body state is acquired in real time. The perceived state information is fed back to the high-level gait and trajectory planning layer, the middle-level whole-body motion optimization layer, and the low-level joint servo control layer, so as to adjust parameters and refresh configurations to achieve the desired control and balance of real-time motion.

8. A storage medium storing a program, characterized in that, When the program is executed by the processor, it implements the method as described in claim 7.

9. A humanoid robot, characterized in that: The humanoid robot gait generation system as described in claim 1 and the real-time motion method as described in claim 7 are employed.

Images