CN-121973208-A - Intelligent planning method and system for autonomous rising action of humanoid robot

CN121973208ACN 121973208 ACN121973208 ACN 121973208ACN-121973208-A

Abstract

A human-shaped robot autonomous rising action intelligent planning method and system belong to the field of robot control and comprise the steps of collecting and fusing internal state and external environment information of a robot in real time to construct a mixed state vector, outputting a high-level action sequence according to the mixed state vector by adopting a deep reinforcement learning strategy network for offline training, converting a behavior semantic instruction into an optimal state track and an optimal control track by adopting a real-time optimizer based on model predictive control, executing a first control quantity of the optimal control track, continuously monitoring an actual state, comparing the actual state with a predicted state of the real-time optimizer based on model predictive control, calculating deviation of the actual state and the predicted state, triggering local correction if the deviation is small, and triggering global rescheduling if the deviation is large. According to the invention, the robot can adaptively cope with various unknown falling postures through the layered intelligent planning architecture, so that a safe, stable and efficient rising track is rapidly generated, and the autonomy and the robustness of the robot are improved.

Inventors

HE YUZHEN
Jiao Dongqiu
QI SHUHENG
HAN YUWEI
LU SIBO

Assignees

北京赛博万智能装备制造有限公司

Dates

Publication Date: 20260505
Application Date: 20260203

Claims (10)

1. The intelligent planning method for the autonomous rising action of the humanoid robot is characterized by comprising the following steps of: step one, acquiring and fusing internal state and external environment information of a robot in real time to construct a mixed state vector S; Step two, a deep reinforcement learning strategy network adopting offline training outputs a high-level action sequence according to the mixed state vector S; Thirdly, converting the behavior semantic instruction into an optimal state track and an optimal control track by adopting a real-time optimizer based on model predictive control; And step four, executing the first control quantity of the optimal control track, continuously monitoring the actual state, comparing the actual state with the predicted state of the real-time optimizer based on model predictive control, calculating the deviation between the actual state and the predicted state, triggering local correction if the deviation is small, and triggering global re-planning if the deviation is large.
2. The intelligent planning method for autonomous rising motion of humanoid robot according to claim 1, wherein in the third step, a barycenter dynamics model is used as a dynamics model, an optimization problem is built and solved, and an optimal state track and an optimal control track in a future time domain are obtained.
3. The intelligent planning method for autonomous rise motion of humanoid robot according to claim 2, wherein the optimization problem is as follows: Wherein H is the future time domain, As a control input for the initial moment of time, As the system state at the current time t, The target reference state is resolved for the behavior semantic instruction a i , For control input, Q and R are weight matrices; for the system state at time t +1, For the system state transfer function, ZMP is the zero moment point, support Polygon is the Support Polygon consisting of the active contact points, For the contact force vector of the ith contact point at time t, Is the ground normal vector of the i-th contact point, For the friction coefficient of the ith contact point, distance is the minimum Distance between the robot and the jth obstacle, robot is the robot, As an obstacle to the j-th of the road, As a safety threshold value, the safety threshold value, In order to control the lower limit value of the input, Is the upper limit value of the control input.
4. The intelligent planning method for autonomous rising motion of humanoid robot according to claim 1, wherein in the fourth step, the deviation δ= lll Sactual-SPREDICTED of the actual state and the predicted state is Sactual in the actual state and SPREDICTED in the predicted state.
5. The intelligent planning method for autonomous rising motion of humanoid robot according to claim 1, wherein in the fourth step, the small deviation is δ+_ε, ε is a deviation threshold, and local correction is triggered at this time, the real-time optimizer based on model predictive control adjusts parameters in the optimization problem and re-optimizes in the next control period to compensate the deviation.
6. The intelligent planning method for autonomous rising motion of humanoid robot according to claim 1, wherein in the fourth step, the large deviation is delta > epsilon, epsilon is a deviation threshold, global re-planning is triggered at this time, the current latest actual state is fed back to the deep reinforcement learning strategy network, a brand new high-level motion sequence is output according to the current latest actual state, and the system starts the bottom-level motion planning again.
7. The intelligent planning method for autonomous rise motion of humanoid robot according to claim 1, wherein in the second step, the deep reinforcement learning strategy network performs large-scale pre-training through reinforcement learning algorithm in highly randomized simulation environment.
8. The autonomous rise motion intelligent planning method of humanoid robots of claim 1, wherein in step two, the reward function of the deep reinforcement learning strategy network is designed to encourage successful standing, minimize energy consumption and time, penalize unstable poses and collisions.
9. The deep reinforcement learning strategy network system applied to the intelligent planning method for autonomous rise motion of humanoid robot according to any one of claims 1 to 8, characterized by comprising: the multi-mode sensing and mixed state construction module is used for collecting and fusing the internal state and external environment information of the robot in real time to construct a mixed state vector S; the high-level strategy decision module adopts a deep reinforcement learning strategy network trained offline to output a high-level action sequence according to the mixed state vector S; the bottom-layer motion planning module converts the behavior semantic instruction into an optimal state track and an optimal control track by adopting a real-time optimizer based on model predictive control; the closed loop execution and self-adaptive adjustment module executes a first control quantity of the optimal control track, continuously monitors the actual state, compares the actual state with the predicted state of the real-time optimizer based on model prediction control, calculates the deviation between the actual state and the predicted state, triggers local correction if the deviation is small, and triggers global rescheduling if the deviation is large.
10. The intelligent planning system for autonomous rise motion of a humanoid robot according to claim 9, wherein the multi-mode sensing and mixing state construction module comprises a body state sensing unit, an environment sensing unit and a state fusion unit, wherein the body state sensing unit comprises an inertial measurement unit, a joint encoder and a force/torque sensor, the inertial measurement unit is used for measuring triaxial acceleration and angular velocity of the robot, the joint encoder is used for measuring the angle of a whole body joint of the robot, and the force/torque sensor is used for measuring the contact state of each end effector with the ground and the contact force and direction of each end effector with the ground.

Description

Intelligent planning method and system for autonomous rising action of humanoid robot Technical Field The invention belongs to the technical field of robot control, and particularly relates to an intelligent planning method and system for autonomous rising actions of a humanoid robot. Background Dynamic balancing and fall recovery of humanoid robots is one of the core challenges in robotics research. At present, the technical path for achieving the rising of the robot is mainly divided into two types, namely a static action sequence based on pre-programming and online track optimization based on a dynamic model. The online track optimization method based on the dynamic model mainly builds a multi-rigid-body dynamic model of the robot, models the lifting process as an optimization problem under various physical constraints (such as joint moment, friction force and balance stability), and utilizes a numerical optimization algorithm (such as Model Predictive Control (MPC)) to solve and generate the motion track online. Such methods rely on the accuracy of the model and are theoretically capable of handling certain disturbances. The closest prior art solution to the present invention is a robot lift method based on-line Model Predictive Control (MPC) which attempts to adapt to different initial conditions by on-line optimization. The specific implementation flow of the scheme is as follows: (1) State sensing, namely acquiring current joint angle and gesture of the robot and contact point information of the robot and the ground through a sensor; (2) Problem modeling, constructing a finite time domain optimization problem based on a simplified robot dynamics model (e.g., centroidal dynamics). The objective function is typically to track a pre-set sequence of lift gestures and minimize energy consumption; (3) And solving the optimization problem in real time according to the current state in each control period to obtain an optimal joint control sequence in a future period. (4) And (3) track execution, namely sending the first group of control instructions obtained by solving to a bottom servo driver for execution. This solution has mainly the following drawbacks: 1. The method has strong dependence on initial conditions and model accuracy, and the scheme is seriously dependent on accurate robot models and environment parameters (such as friction coefficients). When the robot is in an unexpected complex fall pose, the simplified robot dynamics model may be mismatched with the actual dynamics, resulting in an optimization problem with no solution or infeasibility of solutions, and a planning failure. 2. The method has the advantages of high calculation complexity and poor instantaneity, in the scheme, the online optimization calculation amount of the robot whole body dynamics model is huge, the solution time is long, and the high requirement on real-time feedback control in the rising process (the calculation is usually required to be completed in millisecond level) is difficult to meet. This may lead to control delays that unbalance the robot. 3. The scheme is essentially a local optimizer and lacks strategic thinking due to the lack of high-level intelligent decisions and generalization capability. It can only be fine-tuned around a preset rising path, but cannot autonomously decide, like a human being, what to do first and what to do second in the face of a new gesture, such as whether to roll first or support with hands first. Therefore, this approach suffers from poor generalization ability and cannot cope with a large number of unseen fall scenarios. 4. The sensing information is not utilized enough, the robustness is poor, the scheme can not sufficiently integrate the environment sensing information (such as an obstacle) to actively avoid collision, and an effective closed-loop feedback mechanism is usually lacked to cope with uncertainty (such as slipping of a supporting surface) in the execution process, so that the system robustness is not sufficient. Disclosure of Invention The invention provides an intelligent planning method and system for autonomous rising actions of a humanoid robot, which aims to solve the problems of strong dependence on initial conditions and model accuracy, complex calculation, poor instantaneity, weak generalization capability, insufficient robustness and the like in the existing scheme. The invention combines artificial intelligence and model predictive control, and is particularly suitable for autonomous, safe and efficient planning and executing the rising action of the robot from any posture after falling. The technical scheme adopted by the invention for solving the technical problems is as follows: The invention provides an intelligent planning method for autonomous rising actions of a humanoid robot, which mainly comprises the following steps: step one, acquiring and fusing internal state and external environment information of a robot in real time to construct a mixed st