CN-121995754-A - Neural-mechanical coupling software multi-mode motion control method based on pre-training language model

CN121995754ACN 121995754 ACN121995754 ACN 121995754ACN-121995754-A

Abstract

The invention discloses a neural-mechanical coupling software multi-mode motion control method based on a pre-training language model. The system comprises a mass point-spring based soft robot body structure module, a rhythm oscillation signal generation and parameter control module based on a pre-training language model modulation, a muscle driving and mechanical calculation module based on nerve-mechanical mapping and a simulation and debugging module based on physical simulation. The invention is oriented to gait planning and intelligent control of the bionic soft robot in an unstructured environment. Aiming at the problems that gait switching depends on artificial parameters and is difficult to adjust quickly along with instructions, a pre-training language model is introduced, and high-level semantic instructions are mapped into characteristic parameters of the neural controller. The robot can smoothly switch peristaltic, steering and rolling behaviors under the working conditions of a hard plane, viscous fluid and the like through natural language without re-modeling, and a control and analysis method for effectively improving the self-environment adaptability of the software system is provided.

Inventors

ZHENG NENGGAN
LIU YIJIE
TANG WEIHAO
WANG PENGFEI

Assignees

浙江大学

Dates

Publication Date: 20260508
Application Date: 20260127

Claims (9)

1. A neural-mechanical coupling software multi-mode motion control method based on a pre-training language model, which is characterized by comprising the following modules: (S1) a mass point-spring-based soft robot body structure module, wherein a soft robot body executing mechanism with a multi-body section structure is configured according to a real biological structure of drosophila larva, each body section comprises three-dimensional mass point physical characteristics, a distributed muscle executor and a spring connection network, a mass point system comprises head mass points, tail mass points and a plurality of body section mass points which are arranged in a hexagonal topological manner, the muscle executing unit comprises longitudinal, circumferential, oblique and end directional muscle units, and the spring connection network comprises longitudinal springs, oblique springs, annular cross springs and end reinforcing springs; (S2) generating a rhythm oscillation network coupled with excitatory-inhibitory neurons based on a rhythm oscillation signal modulated by a pre-training language model, establishing a rhythm oscillation network coupled with excitatory-inhibitory neurons, receiving an external natural language instruction by each oscillation node corresponding to one body node of the robot, analyzing and generating a control token through the pre-training language model, and further adjusting parameters of the CPG rhythm oscillator in real time to generate a voltage pulse/rhythm driving signal with body phase delay; (S3) a muscle driving and mechanical calculating module based on nerve-mechanical mapping converts the rhythm oscillation signals into activation instructions of different types of muscle execution units, calculates the active tension of the muscles by combining the mechanical behavior characteristics of the muscles, and performs various physical actions such as spring force, damping force, friction force, adhesion force, collision force, volume constraint force and the like to complete the dynamic solution of a texture system; And S4, based on a simulation and debugging module of physical simulation, invoking a physical engine to input the activation state of the muscle execution unit into a dynamic solver, wherein the physical constraint force is used for maintaining the stability of body deformation, and controlling the soft robot to complete a target movement mode under a specified working condition.
2. The method according to claim 1, wherein in the module (S1), the physical skeleton of the soft robot body is arranged step by taking the axial coordinates of the body segments as a reference, and the physical boundary is defined by setting the side length of the hexagon by the radius of the section, so as to ensure that the body segments maintain stable mechanical topological constraints under different deformation modes such as longitudinal compression, circumferential contraction and oblique bending, and the hexagonal particles are connected by springs, so as to limit the relative displacement of the local particles and allow continuous smooth morphological evolution under large deformation.
3. The method according to claim 1, characterized in that in the module (S1), the soft robot comprises an actuator array consisting of groups of artificial muscles, the muscle types and the spatial orientations are mapped using a unique addressing coding protocol, firstly the muscles are divided into four classes of longitudinal, circumferential, oblique and end orientations according to their structural direction, secondly the unique numbers are determined according to their spatial orientations in the cross-section or body-segment axial coordinate system, and finally global indexes are formed in combination with the body segment numbers, by means of which the control system can drive the individual rhythms of the specific execution units.
4. The method according to claim 1, wherein in block (S2), the rhythmic oscillation signal generation block comprises a rhythmic signal generator that generates a rhythmic driving waveform using a neuromechanical control logic: Wherein: i is the body segment number of the body, Excitation, inhibition, and sensory feedback variables of the ith segment, respectively; a,b,c,d,e,f, As coupling coefficients, respectively controlling the excitation-inhibition connection strength of the current segment and the adjacent segment, the influence degree of sensory feedback and the relative weight of signal transmission between the adjacent segments; For a gain parameter, for adjusting the maximum activity level of the excitation and suppression unit; As a time constant, controlling the response speed of each unit; For external pulse input, only periodically triggering at the head body node for starting forward propagation of rhythmic waveforms; a nonlinear transfer function for the cell is used to define the node output range and maintain periodicity; The generator outputs voltage signals/driving signals with phase delay according to the body section number, and carries out real-time motion phase compensation by virtue of a sensory feedback variable, the control logic enables the excitation signals of adjacent sections to keep fixed delay in a time domain, a traveling wave mode from back to front is formed on the whole body, and the periodicity comes from the limit cycle characteristic of a first-section periodic pulse triggering and nonlinear dynamics equation, so that continuous and stable rhythm generation is realized under the combined action of the two.
5. The method according to claim 1, wherein the generation and conversion of the multiple motion modes in the module (S2) are based on the generation and adjustment of a rhythm oscillation control parameter of a pre-training language model, and the parameter modulation process specifically includes that a semantic command analysis unit in a control system receives an externally input natural language command, performs semantic feature extraction and decoding on the command by using the pre-training language model, generates a set of control tokens corresponding to a specific motion mode, including a rhythm feature parameter, a phase delay coefficient and a driving amplitude, and is used for configuring a bottom layer parameter of a rhythm signal generator in real time so as to drive a software robot to perform gait switching and mode changing between different working conditions; In a peristaltic advancing mode, the system sequentially triggers excitation signals in muscle groups of body joints to form contraction waves which are transmitted from tail to head along the axial direction, longitudinal muscle activation values of all the body joints have certain phase delay, so that the contraction waves are stably transmitted and push the whole body to move forwards, in a swinging and turning mode, the controller stably contracts the longitudinal muscles at the bending direction side to provide steering force of the body through coordinating different activation modes of the muscles at the left side and the right side, the vibration signals are provided at the other side to generate periodic fluctuation, the robot is pushed to stably turn towards a new direction, in a rolling mode, the nerve vibration signals are integrally redistributed to oblique muscle groups of the whole body and form coupling driving with longitudinal and circumferential muscles, the coupling driving is embodied in that all the oblique muscles are alternately activated to form spiral contraction waves along the longitudinal axis, meanwhile, the longitudinal and circumferential muscles provide structural support and section contraction, continuity and stability of the rolling process are guaranteed, and the cooperative effect enables the robot to rapidly complete integral overturning, and emergency avoidance actions similar to the escaping actions of fruit larvae.
6. The method according to claim 1, wherein in the module (S3), the rhythmic oscillation electric signal generated by the rhythmic signal generator in the module (S2) is obtained, and the electric signal intensity is converted into an activation driving command corresponding to the artificial muscle execution unit according to a preset neuro-mechanical mapping protocol, and the driving command is then combined with a length-tension physical characteristic curve and a speed-tension physical characteristic curve of the software robot body to output physical tension control values acting on each group of muscle units in real time To overcome environmental resistance and generate power for driving the robot to displace: . Wherein the method comprises the steps of Is the maximum contractile force of the muscle unit; muscle activation values obtained for the neuro-mechanical mapping, derived from the output of the rhythmic oscillating signal, As a function of length-tension relationship, for describing the tension variation of a muscle at different elongation lengths; the influence of the muscle contraction speed on the mechanical output is reflected as a speed-tension relation function; the term corresponds to the parallel spring force A spring rate coefficient), the passive elasticity of the simulated muscle and its ancillary tissues, ensuring that a restoring force is still provided when not activated; Is a damping force term Damping coefficient), exhibits viscous drag effects, is used to suppress excessively rapid length changes, avoids numerical instability, For the amount of muscle extension, Is the shrinkage rate.
7. The method according to claim 1, wherein in the module (S4), the simulation and debugging module provides real-time motion monitoring, records motion trajectories and performance indexes of the soft robot under different parameter combinations, specifically, in each simulation time step, solves the speed and position of updated particles according to the resultant action of various external forces, ensures the continuity and stability of the system state evolving along time, rapidly evaluates the effects of different control parameters in a unified simulation platform by the module, realizes the verification of multiple motion modes, and provides reliable theoretical basis for subsequent hardware design and optimization.
8. The method of claim 7, wherein the various external forces include one or more of muscle contraction force, spring force, damping force, friction force, adhesion force, volume restraining force, and crash reaction force.
9. The method of claim 1 for motion gait planning and intelligent control of a biomimetic soft robot in an unstructured environment.

Description

Neural-mechanical coupling software multi-mode motion control method based on pre-training language model Technical Field The invention relates to the technical field of software robot control, in particular to a neural-mechanical coupling software multi-mode motion control method based on a pre-training language model. Background The soft bionic robot is an intelligent robot system which takes flexible materials as a main body structure and is inspired by invertebrates in the nature. Compared with the traditional rigid robot, the soft robot has the structural characteristics of high flexibility and continuous deformability, can complete tasks in complex, narrow or unstructured environments, and becomes a research hotspot in the fields of bionics and intelligent control. However, due to the high degree of coupling between the morphology and behavior of soft robots, there are still implementation difficulties in motion control, mechanical modeling, and multi-motion pattern generation. On the control level, the current international mainstream research has attempted to realize high-level task planning of robots by utilizing the powerful semantic parsing capability of a pre-training language model (LLM). However, for the discrete tasks of the rigid robot, how to convert the abstract language instruction into accurate continuous physical driving parameters (such as muscle activation degree and body joint phase delay) for the soft robot with infinite freedom degree characteristics, an effective end-to-end mapping mechanism is not yet available, so that serious technical faults exist between instruction analysis and physical execution. In the aspect of a motion execution mechanism, although a neural control method based on a Central Pattern Generator (CPG) can generate stable bionic rhythmic motion, the existing achievements depend on a manually preset parameter mapping table or a specific gait library, so that the soft robot lacks flexibility and cross-mode self-adaption capability when facing multi-mode switching such as peristaltic motion, steering and rolling. In the aspect of dynamic modeling and feedback, although simplified algorithms such as particle spring models and the like improve the calculation efficiency, the control law and mechanical response are often treated separately in the prior art, and the deep coupling effect of nerve-mechanics is ignored, so that the driving system of the soft robot is difficult to compensate the structural stability problem generated by large deformation in real time when the soft robot performs fine movements, and the operation precision and physical interaction capability of the soft robot in a dynamic complex environment are limited. In summary, developing a software multi-mode motion control method that can integrate semantic intelligence of a pre-training language model, CPG rhythm characteristics and multi-body mechanical behaviors has become a key to breaking through the bottleneck of autonomous operation of a software robot. Disclosure of Invention The invention aims to provide a neural-mechanical coupling software multi-mode motion control method based on a pre-training language model, which is used for solving the problems of independent modeling of structure and control, inflexibility in multi-motion mode switching, dependence on manual setting of parameter configuration and the like of the traditional software robot. The method takes multi-mode motion of drosophila larvae as biological elicitation, establishes a multi-body-segment soft robot structural model, a rhythm oscillation control network and a neural-mechanical mapping relation, generates rhythm control parameters by utilizing a pre-training language model, and realizes automatic scheduling of the motion mode of the multi-body-segment soft robot. The specific implementation and execution flow of the method of the invention are as follows: (S1) constructing a mass point-spring-based soft robot body structure module, and establishing a physical topological frame formed by a three-dimensional mass point system, an elastic connection network and a distributed artificial muscle execution unit, wherein the module is used as an execution main body of the soft robot, and the execution mechanism is ensured to have multi-degree-of-freedom flexible deformation and physical adaptability under biological inspiring through spatial layout configuration of the muscle units. And S2, establishing a rhythm oscillation signal generation and parameter control module based on the pre-training language model modulation, integrating a pre-training language model interface for decoding a received natural language instruction into a control token, dynamically adjusting parameters of a built-in excitability-inhibitory neuron coupling network by the token, generating a bottom layer rhythm driving signal with space-time phase delay, and realizing intelligent modulation of high-level semantics on a bottom layer movement mode. And