CN-122021709-A - Human-shaped robot action redirection optimization method based on physical constraint

CN122021709ACN 122021709 ACN122021709 ACN 122021709ACN-122021709-A

Abstract

The invention discloses a human-shaped robot action redirection optimization method based on physical constraint, and belongs to the field of human-shaped robot imitation learning. The existing action redirection technology is not fully considering the dynamic constraint and physical characteristics of a robot, so that the problems of non-physical floating, ground penetration and insufficient dynamic stability are easy to occur when the high dynamic action is migrated. According to the invention, human motion data are acquired through motion capture and converted into parameterized parameters, three types of physical constraint modules including motion stability, body height and joint motion range are introduced in the redirection process, constraint loss and task reconstruction loss are optimized in a cooperative manner by adopting a gradient descent algorithm, and a robot joint track conforming to a physical rule is output. The invention can eliminate the non-physical movement phenomenon, improve the dynamic stability and the physical rationality of the movement, provide guarantee for accurate migration of the high dynamic movement, and is suitable for simulating a learning scene of the humanoid robot.

Inventors

SUN TAO
SHAO DONGHAO
Dai Menglin
YU CHAO

Assignees

鹿明机器人科技(深圳)有限公司
鹿明科技(苏州)有限公司

Dates

Publication Date: 20260512
Application Date: 20251202

Claims (10)

1. The human-shaped robot action redirection optimization method based on physical constraint is characterized by comprising the following steps of: S1, acquiring human body complex motion data, and converting the human body complex motion data into parameterizable motion key parameters through human body model calculation, wherein the motion key parameters at least comprise a human body motion follow-up coordinate system and the spatial position and posture of each joint; s2, converting the parameterized motion key parameters into initial motion parameters suitable for a robot model through a redirection process, and generating an initial motion track of a robot joint space; s3, three types of physical constraint modules are introduced in the redirection process, wherein the three types of physical constraint modules are a motion stability constraint module, a body height constraint module and an articulation motion range constraint module respectively; S4, adopting a gradient descent algorithm, carrying out cooperative optimization on constraint losses and task reconstruction losses corresponding to the three types of physical constraint modules, adjusting the initial motion track, and outputting a target motion track conforming to the dynamic characteristics and the physical laws of the robot.
2. The humanoid robot motion redirection optimization method based on physical constraints of claim 1, wherein the "constraint loss calculation of the motion stability constraint module" in step S3 includes: S311, calculating the overall mass center of the robot, namely carrying out weighted average on the mass center mass of each connecting rod of the robot and the space position of the mass center of each connecting rod to obtain the overall mass center; S312, calculating a robot plantar pressure center, namely generating a differentiable plantar support point cloud tensor based on plantar connecting rod positions and geometric parameters, calculating vertical distances between each point in the point cloud tensor and the ground to distinguish an invasive ground support point and a ground support point, setting support force calculation rules for the two types of support points respectively, and weighting and averaging the positions of the support points according to corresponding support forces to obtain the plantar pressure center; S313, calculating the motion stability constraint loss, namely calculating the Euclidean distance between the projection of the integral mass center on the XOY plane and the plantar pressure center, and taking the Euclidean distance as the motion stability constraint loss.
3. The method according to claim 2, wherein the step S312 of "rule for calculating supporting force" includes setting supporting force to be in a linear increasing relationship with the distance of the supporting point from the ground, the supporting force being calculated by a formula W pen ＝1-k pen ·cop pen , wherein k pen is a scaling factor of the supporting force from the ground, cop pen is the distance from the supporting point from the ground, setting supporting force to be in an exponential decreasing relationship with the distance from the supporting point from the ground, and the supporting force being calculated by a formula W float ＝exp(-k float ·cop float ), wherein k float is a scaling factor of the supporting force from the ground, and cop float is the distance from the supporting point from the ground.
4. The method for redirecting and optimizing human-shaped robot actions based on physical constraints according to claim 1, wherein the step S3 of calculating the constraint loss of the body height constraint module comprises calculating the difference between the body height and the initial body height at each moment in the motion process of the robot, taking the square of the difference or the absolute value of the difference as the body height constraint loss, or carrying out average processing on the difference at all moments, and taking the square of the average difference or the absolute value of the average difference as the body height constraint loss.
5. The method for redirecting and optimizing human-shaped robot actions based on physical constraints according to claim 4, wherein the step S3 of calculating constraint losses of the constraint module of the joint motion range comprises the steps of presetting a lower angle limit and an upper angle limit of each driving joint of the robot, performing out-of-range inspection on actual angles of each driving joint at each moment, applying secondary punishment to the excess part if the actual angles exceed the lower angle limit or the upper angle limit, and summing punishment values of all joints to obtain constraint losses of the joint motion range.
6. The method for redirecting and optimizing human-shaped robot actions based on physical constraints according to claim 1, wherein the "collaborative optimization" in step S4 is implemented by using a multi-objective loss function, and the expression of the multi-objective loss function is L task is task reconstruction loss for maintaining original human motion characteristics, L 1 is motion stability constraint loss, L 2 is body height constraint loss, L 3 is joint motion range constraint loss, alpha 1 、α 2 、α 3 is a weight coefficient corresponding to L 1 、L 2 、L 3 respectively, and alpha 1 、α 2 、α 3 is a positive number.
7. The method for redirecting and optimizing human-shaped robot actions based on physical constraints according to claim 2, wherein "plantar support point cloud tensor" in step S312 is generated by the formula: Pad ijk ＝T foot ·[x i ,y j ,z k ,1] T , wherein x min 、x max is the minimum value and the maximum value of the foot geometric envelope frame in the x-axis direction, N x is the point cloud resolution in the x-axis direction, y min 、y max is the minimum value and the maximum value of the foot geometric envelope frame in the y-axis direction, N y is the point cloud resolution in the y-axis direction, z min 、z max is the minimum value and the maximum value of the foot geometric envelope frame in the z-axis direction, N z is the point cloud resolution in the z-axis direction, and T foot is a homogeneous transformation matrix of foot connecting rods for transforming local coordinate points [ x i ,y j ,z k ,1] T ] to a world coordinate system.
8. The method for redirecting optimization of humanoid robot actions based on physical constraints of claim 5, wherein the "quadratic penalty" is calculated by the formula Calculating, wherein T is the total number of motion time frames, N dof is the number of robot driving joints, q i (T) is the actual angle of the ith joint in the T frame, q i,max is the upper limit of the angle of the ith joint, and q i,min is the lower limit of the angle of the ith joint.
9. The method according to claim 5, wherein the "complex motion data of human body" in step S1 is obtained by a motion capture device, and the motion capture device comprises an optical motion capture device, an inertial motion capture device, or an electromagnetic motion capture device.
10. The method for redirecting and optimizing motion of a humanoid robot based on physical constraints according to claim 1, wherein the "physical rationality" of the motion trajectory of the target is verified by motion smoothness, the motion smoothness is measured by calculating the sum of norms of accelerations of each joint of the robot, and the smaller the sum of the norms of the accelerations is, the higher the motion smoothness is.

Description

Human-shaped robot action redirection optimization method based on physical constraint Technical Field The invention relates to the technical field of humanoid robot imitation learning, in particular to a humanoid robot action redirection optimization method based on physical constraint. Background The human-shaped robot realizes the transfer of human action skills through imitative learning, which is one of the core directions in the robot autonomous learning field, wherein the action redirection technology is used as a key link of imitative learning, and the human action track is corrected and parameterized and mapped to the robot joint space, so that the robot learning cost can be greatly reduced, the safety of the deployment of the motion strategies is improved, and the robot can realize the efficient action transfer only by a small amount of human demonstration samples. However, the prior action redirection technology has the obvious technical defect that the dynamic constraint (such as joint motion limit and mass center balance rule) and physical characteristic limitation (such as interaction rule between foot end and ground) of the robot body are not fully considered, so that the motion trail which is against the physical rule is easy to generate when the high dynamic characteristic actions such as migration running, jumping, rapid steering and the like are performed. The concrete steps are as follows: Non-physical floating, namely, the height of the robot body deviates from a reasonable range in the motion process of the robot, and the phenomenon of suspension or non-supporting rising occurs; ground penetration, namely, the foot end of the robot excessively invades a ground model, and the physical common sense of 'ground impenetrability' is violated; The dynamic stability is insufficient, namely, the robot is easy to topple over due to the fact that the motion is completed only by means of limited supporting points of feet and the mass center projection exceeds the supporting area. The defects seriously restrict the application of the action redirection technology in complex dynamic environments (such as home service and industrial collaboration), and an optimization method capable of fusing physical constraints and guaranteeing the action rationality is needed. Disclosure of Invention The invention aims to provide a humanoid robot action redirection optimization method based on physical constraint, which aims to solve the two core problems of the existing non-physical motion phenomenon and insufficient dynamic stability. The technical scheme for solving the technical problems is as follows: A humanoid robot action redirection optimization method based on physical constraint comprises the following steps of S1, obtaining human body complex motion data, converting the human body complex motion data into parameterizable motion key parameters through human body model calculation, wherein the motion key parameters at least comprise a human body motion follow-up coordinate system and space position postures of joints, S2, converting the parameterized motion key parameters into initial motion parameters suitable for a robot model through a redirection process to generate an initial motion track of a robot joint space, S3, introducing three types of physical constraint modules in the redirection process, wherein the three types of physical constraint modules are a motion stability constraint module, a body height constraint module and a joint motion range constraint module respectively, S4, adopting a gradient descent algorithm, performing collaborative optimization on constraint losses corresponding to the three types of physical constraint modules and task reconstruction losses, adjusting the initial motion track, and outputting a target motion track conforming to the dynamic characteristics and the physical laws of the robot. The step S3 is characterized in that constraint loss calculation of the motion stability constraint module comprises the steps of S311 calculating a robot integral mass center, S312 calculating a sole pressure center of the robot, wherein the sole pressure center is obtained by carrying out weighted average on mass center mass of each connecting rod of the robot and space positions of mass centers of each connecting rod, S312 calculating a differential sole support point cloud tensor based on sole connecting rod positions and geometric parameters, calculating vertical distances between each point in the point cloud tensor and the ground to distinguish an invasive ground support point and a floating ground support point, setting a support force calculation rule for the two support points respectively, and carrying out weighted average on positions of each support point according to corresponding support force to obtain the sole pressure center, S313 calculating motion stability constraint loss, wherein the Euclidean distance between a projection of the integral mass center on an XOY