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

Abstract

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

Description

Technical Field

[0001] The invention relates to the field of soft robot control technology, and in particular to a neural-mechanical coupled soft multi-mode motion control method based on a pre-trained language model. Background Technology

[0002] Soft biomimetic robots are a class of intelligent robotic systems inspired by invertebrates in nature, with flexible materials as their main structural component. Compared to traditional rigid robots, soft robots possess high compliance and continuously deformable structural characteristics, enabling them to complete tasks in complex, confined, or unstructured environments, making them a research hotspot in the fields of biomimetic and intelligent control. However, due to the high coupling between the morphology and behavior of soft robots, there are still challenges in achieving motion control, mechanical modeling, and multi-motion mode generation.

[0003] At the control level, current mainstream international research has attempted to leverage the powerful semantic parsing capabilities of pre-trained language models (LLMs) to achieve high-level task planning for robots. However, these methods are mainly aimed at discrete tasks for rigid-body robots. For soft robots with infinite degrees of freedom, there is currently a lack of effective end-to-end mapping mechanisms for converting abstract language instructions into precise continuous physical driving parameters (such as muscle activation levels and segment phase delays), resulting in a serious technical gap between instruction parsing and physical execution.

[0004] At the level of motion execution mechanism, although neural control methods based on central pattern generators (CPG) can generate stable biomimetic rhythmic movements, existing results mostly rely on artificially preset parameter mapping tables or specific gait libraries, resulting in soft robots lacking flexibility and cross-modal adaptive capabilities when facing multi-mode switching such as crawling, turning, and rolling.

[0005] At the level of dynamic modeling and feedback, although simplified algorithms such as the point mass spring model improve computational efficiency, existing technologies often process the control law and mechanical response separately, ignoring the deep coupling effect of "neuromechanics". This makes it difficult for the drive system of soft robots to compensate for the structural stability problems caused by large deformation in real time when performing fine movements, which limits the operation accuracy and physical interaction capability of software robots in dynamic and complex environments.

[0006] In summary, developing a soft multi-mode motion control method that integrates the semantic intelligence of pre-trained language models, CPG rhythmic features, and multi-segment mechanical behavior has become crucial for overcoming the bottleneck of autonomous operation of soft robots. Summary of the Invention

[0007] The purpose of this invention is to provide a neuro-mechanical coupled multi-modal motion control method for soft robots based on a pre-trained language model, addressing problems such as independent modeling of structure and control, inflexible switching between multiple motion modes, and reliance on manual parameter configuration in existing soft robots. Inspired by the multimodal motion of fruit fly larvae, this invention establishes a structural model of a multi-segment soft robot, a rhythmic oscillation control network, and a neuro-mechanical mapping relationship. It then utilizes a pre-trained language model to generate rhythmic control parameters, enabling automatic scheduling of the multi-segment soft robot's motion modes.

[0008] The specific implementation and execution flow of the method of this invention are as follows: (S1) Construct a soft robot body structure module based on mass-spring, and establish a physical topology framework consisting of a three-dimensional mass system, an elastic connection network and distributed artificial muscle execution units. This module serves as the execution body of the soft robot. Through the spatial layout configuration of the muscle units, it ensures that the execution mechanism has bio-inspired multi-degree-of-freedom flexible deformation and physical adaptability.

[0009] (S2) Establish a rhythmic oscillation signal generation and parameter control module based on pre-trained language model modulation, integrate the pre-trained language model interface, and use it to decode the received natural language instructions into control tokens; use the tokens to dynamically adjust the parameters of the built-in excitatory-inhibitory neuron coupling network to generate a low-level rhythmic driving signal with spatiotemporal phase delay, thereby realizing intelligent modulation of the low-level motion modality by high-level semantics.

[0010] (S3) Implement a drive control module based on neuro-mechanical mapping, which converts the rhythmic drive signal into specific tension control commands acting on each artificial muscle actuator according to a physical coding protocol; the control module drives the robot actuator to complete the physical coordinated deformation of each segment by adjusting the activation sequence of different types of muscles.

[0011] (S4) Construct a simulation and debugging module based on physical simulation, use physical dynamic monitoring mechanism to obtain mass state data of actuator during motion, and use dynamic feedback to verify the motion stability of the robot in real time; this module is used to verify the consistency between physical execution process and instructions, and to perform closed-loop optimization and hardware debugging of control parameters.

[0012] Furthermore, the specific content of the soft robot body structure module based on mass-spring constructed in (S1) is as follows: The soft robot body of this invention adopts a physically modular layered architecture, consisting of three parts: a three-dimensional mass system, a muscle execution system, and a spring connection network.

[0013] (1-1) The mass system described herein is based on the morphological data of fruit fly larvae to construct the basic framework of the robot body. First, the positions of the key control points at the head and tail are determined, and a structural framework consisting of 62 physical nodes is configured in three-dimensional space. This framework includes one key fixed point at the head and tail and 10 functional segments. Each segment is arranged in a regular hexagonal topology by 6 nodes. By controlling the cross-sectional radius parameters of the nodes, the physical outline dimensions and continuous deformation space of the robot are defined, providing mechanical structural support for multi-mode motion.

[0014] (1-2) The muscle execution system described herein arranges four types of functional muscle execution units on a mass frame, including end muscles (6 sets of radial connections at the head and tail to ensure precise drive and posture stability), longitudinal muscles (connecting adjacent body segments to vertices in the same orientation to achieve axial extension and contraction), circumferential muscles (circularly connecting adjacent vertices in the same segment to control cross-sectional contraction), and oblique muscles (diagonally connecting specific vertices of adjacent body segments to provide torsional and asymmetric bending capabilities). All muscle execution units are managed by body segment group and a “type-orientation-body segment number” coding system (e.g., long_right_seg0) is established to identify and independently address each muscle unit, accurately mapping control signals to the specific physical location of the execution unit.

[0015] (1-3) The spring connection network is used to establish the mechanical structural constraints of the robot body, maintain the overall topological stability, and control the relative motion between body segments. The network forms an end-strengthening structure through radial connections, and multiple sets of compression- and shear-resistant elastic support elements (including longitudinal, oblique, and ring-shaped cross supports) are arranged inside and between body segments. All support elements are topologically deduplicated and arranged in an orderly manner to form a non-redundant physical connection network, ensuring that the actuator has continuous mechanical stiffness and motion stability in complex motion responses.

[0016] Furthermore, the specific content of the (S2) rhythmic oscillation signal generation and parameter control module based on pre-trained language model modulation is as follows: (2-1) This module employs a distributed excitatory-inhibitory neuron coupled rhythmic oscillation network. The network assigns an independent rhythmic oscillation control node to each segment of the soft robot, controlled by excitatory units. Inhibitory units and sensory feedback unit The structure consists of all control nodes connected sequentially along somatic segments, forming a chain-like control structure. The state evolution within each control node satisfies the following set of continuous-time neurodynamic equations: in: i Number the controlled body segments. They are respectively the first iSegmental excitation, inhibition, and sensory feedback variables; a, b, c, d, e, f, The coupling gain coefficient controls the excitation-inhibition connection strength and the degree of sensory feedback influence between the control nodes of this segment and the adjacent segment, thereby determining the coordination of gait. This is the gain parameter; It is a time constant, used to control the response speed of each unit; An external pulse input signal is periodically triggered by the main controller in the head and sac to initiate the waveform propagation of the whole-body gait. The nonlinear modulation function of the control unit performs nonlinear shaping on the output drive waveform (this implementation uses ReLU, but it can also be replaced with Sigmoid to obtain different waveform characteristics).

[0017] (2-2) To ensure that the rhythm signal is updated stably over time, this module uses a high-precision numerical solution algorithm followed by Euler integration to solve the node state variables in real time. After the node output has gone through several update cycles, the system will automatically perform amplitude correction based on the node output amplitude to prevent numerical divergence or signal drift, and ensure that the output amplitude of each actuator is stable and the phase relationship is constant, resulting in a rhythmic dynamic wave.

[0018] (2-3) This module also introduces a pre-trained language model to parse natural language input and automatically generate motion control parameters (control tokens) that adapt to different task requirements. These parameters include the phase delay between segments in the rhythmic oscillation signal, activation amplitude, muscle distribution pattern, etc., to drive the robot to perform different motion modes, such as peristaltic forward movement, head-turning, and rolling.

[0019] At the instruction parsing level, a language processor fine-tuned for the task instructions is used as the core decoding module. By constructing a dedicated dataset containing natural language task instructions and corresponding hardware control parameters, the processor gains the ability to understand task instructions in the robot control domain. During the system optimization phase, the internal mapping logic is optimized by minimizing the control parameter prediction error. Its control parameter optimization loss function not only considers the prediction error but also introduces a task-related weighting term, ensuring that the model can generate control parameters more accurately for each motion mode. , . in, These are the control parameters for model prediction. These are the actual control parameters. N It is the sample size. M It is the number of control parameters for each sample. These are the weighting coefficients for each task. The prediction errors for each individual task are summed together to obtain the total error. For example, the wriggling pattern is weighted more highly, while the head-shaking pattern is weighted less highly, to optimize the model's performance on specific tasks. The Adam optimizer is used for training, updating the model's parameters by minimizing the loss function.

[0020] The fine-tuned language model can process natural language task instructions and transform them into robot control parameters. This process consists of two stages: In the language encoding stage, the input natural language instructions first undergo tokenization and word vector mapping. The model uses a self-attention mechanism to calculate the contextual relationships of each word, thereby understanding the important components of the instruction. For example, the instruction "creep forward" requires identifying "creep" as the movement type and "forward" as the target direction. The Transformer architecture's self-attention mechanism calculates the relationships between words using the following formula: . in, It is a query matrix. It is a key matrix. It is a value matrix. This refers to the dimension of the key matrix. ()express This function allows the model to dynamically adjust the weight of each word, resulting in a more accurate contextual representation. After multiple layers of Transformer encoding, the input text is transformed into a high-dimensional semantic vector. It contains contextual information about the input task instructions.

[0021] In the task intent parsing phase, the model is based on semantic vectors. Contextual reasoning is performed to extract the parameters required for the control task. These include target motion patterns, muscle activation amplitude, phase delay, etc. The contextual reasoning network extracts task-related structured control parameters from these semantic vectors, ensuring that the generated control signals meet the actual task requirements. The processing logic of the intent extraction network is as follows: . in, It is a context-based reasoning network that generates control parameters by learning the relationship between the task and the control parameters. .

[0022] After task intent parsing, the model generates control tokens based on the parsed control parameters. These control tokens are the specific control signals required to execute the robot's task. The control token generation process is based on an autoregressive decoder. In other words, the generation of each control token depends on the output of the previous control token and the current task intent: in, It is the previous control token. It is the current semantic vector. This is the current environmental feedback. The autoregressive decoder ensures the robot can accurately execute movements according to task instructions by progressively generating control tokens. Each control token contains multiple control parameters, such as the phase delay between segments (…). Δ The control tokens (Φ), activation amplitude (α), and muscle distribution pattern (M) are transmitted to the robot's execution system, driving the robot to execute corresponding motion patterns according to different tasks. For example, in the "peristaltic forward" mode, the control tokens adjust the phase delay and activation amplitude of longitudinal muscle contractions between segments to form a stable propulsive contraction wave. For the "head-turning" mode, the control tokens adjust the activation amplitude of the left and right segments and coordinate the activation patterns of the muscles on both sides to achieve directional deflection. For the tumbling mode, the control tokens adjust the activation and phase delay of the oblique muscles to generate a spiral contraction wave along the longitudinal axis and coordinate the coupling effect of the oblique and longitudinal muscles to ensure the stability of the tumbling process.

[0023] The intelligent drive architecture built in this module, which follows the entire path from "instruction → semantics → control parameters → control token → action execution" and is combined with continuously received sensory feedback signals, ensures the motion stability, decision interpretability, and physical adaptability of the soft robot in dynamic environments.

[0024] Furthermore, the specific content of the drive control module based on neural-mechanical mapping implemented in (S3) is as follows: (3-1) This module receives the output of module (S2) and maps the oscillation signal to the change in the target length of the spring. The process is as follows: the controller first receives the normalized excitation signal output from the rhythmic oscillation module, which contains bilateral independent signals from 12 individual segments, and then uses... (Left-side signal) and (The signal on the right) indicates that, i ∈[0,11] represents the somatic segment index; according to the physical arrangement and addressing protocol of the artificial muscle actuator, the signals are precisely allocated to the corresponding execution units: head muscles (i = 0) receive head somatic segment signals, somatic muscles (i = 1~10) receive signals from the corresponding intermediate somatic segments, and tail muscles (i = 11) receive tail somatic segment signals. Furthermore, the signals are received according to the left and right sides of the somatic muscles. and This enables the mapping of multi-channel rhythmic signals to the drive of the entire body's artificial muscle actuation array.

[0025] (3-2) After obtaining the neural excitation value of each muscle, it is converted into active tension acting on the physical structure by calculation based on the mechanical behavior characteristics of the muscle. The specific process is as follows: First, the input neural excitation value is smoothed by a first-order activation dynamics model to simulate the gradual process of the muscle from an excited state to a mechanically activated state: . in This is a nerve excitation signal. To activate the kinetic coefficients, Real-time activation state variables for muscle actuators.

[0026] Then, for each muscle, based on the position vectors of its two anchored points... Calculate current muscle length and relative velocity .

[0027] In the force generation process, the physical and mechanical response characteristics of artificial muscles are comprehensively considered, including the force-length relationship. Force-velocity relationship () ) and passive elastic elements ( ) contributions.

[0028] The force-length relationship is described by an exponential function to represent the peak force characteristics near the optimal length. Indicates the optimal length for muscle exertion. , Preset parameters: . The force-velocity relationship is calculated separately for the two states of shortening and lengthening: During shortening: When stretched: ,in , These are preset parameters. This represents the maximum muscle contraction speed.

[0029] Passive elastic element Used to describe the additional passive contribution beyond the optimal length.

[0030] Ultimately, muscle tension The calculation formula is: . in This represents the maximum tension coefficient of the muscle.

[0031] The calculated tension is along the muscle direction (from the point mass). a Pointing to the mass b unit vector When applied to the two corresponding particles, they form a pair of active driving forces that are equal in magnitude and opposite in direction: . Through the aforementioned closed-loop driving process, the system achieves precise conversion from underlying rhythmic electrical signals to actual mechanical tension, driving the software actuator to generate motion deformation that matches the command, thus completing the dynamic physical control of the software structure.

[0032] Furthermore, the specific content of (S4) constructing the simulation and debugging module based on physical simulation is as follows: (4-1) This module builds a simulation engine based on the spring-mass dynamics model to achieve joint optimization of structural physical parameters and control parameters, and evaluates the motion efficiency and physical stability of the soft robot in real time under different modes in digital space, providing accurate parameter indicators for the manufacturing and driving strategies of physical equipment. The specific simulation and debugging process includes mechanical solution, motion state update and performance optimization evaluation.

[0033] (4-2) In module (S1), each mass point in the soft robot body is connected by springs to form a flexible structure. The position of each mass point in three-dimensional space is set as... The speed is ,in i This provides an index for point masses. The mechanical behavior between point masses is mainly influenced by spring force, gravity, muscle tension, adhesion, friction, and collision force. The system monitors the physical state of each point mass in real time and calculates the values ​​for each mass. i The net external force received Specifically, it includes: Structural internal forces: based on the restoring force generated by the spring connection network, and the active muscle tension output by the (S3) module. Environmental interaction forces: frictional force between the mass and the working plane, collision recovery force to avoid penetrating the plane, and exponentially decaying ground adhesion force simulating the characteristics of soft biological organisms; system damping and gravity: linear damping force is introduced to suppress numerical oscillations, and the gravity contribution is calculated in combination with the mass of the mass.

[0034] To maintain the volumetric stability of the soft robot during deformation, volume constraint compensation is introduced. First, the system calculates the closed shell surface... The directed volume, real-time monitoring of the instantaneous volume of the body. The calculation process utilizes the geometric centers of the nodes. Perform coordinate normalization and then perform piecewise volume accumulation based on the normal direction of the surface triangular mesh: . After the volume calculation is completed, the volume at the current time step is obtained. While retaining the volume of the previous time step To estimate the rate of volume change .

[0035] Next, volumetric constraint forces are generated. This invention employs a combination of adaptive stiffness and damping to achieve a balance between flexibility and stability: , . in It is the volume of the soft robot when it is at rest. The preset basic volumetric stiffness coefficient, This represents the real-time effective stiffness after nonlinear modulation. Let be the volume damping coefficient, and finally be the global correction value. This scalar correction value then needs to be distributed to each vertex to form forces. The system accumulates the values ​​for each vertex using a pre-computed vertex-triangle adjacency map. i Local volume gradient direction .when (Under Implementation) When ), normalize it to a direction. To ensure global conservation and avoid excessive forces at a single point, the code corrects the scalar deviation. Distribute evenly to From the vertices, the final volume constraint force is obtained. : . That is, in the direction of local volume increase ( Apply and Same number, and Force with opposite sign — when the current volume is greater than the reference volume ( )hour, The direction of the force causes the surface to contract inward; when the volume is smaller than the reference value, the direction of the force promotes outward expansion; while the damping term... This suppresses volume oscillations and improves numerical stability. In practice, the force is accumulated atomically onto the particle, and the current volume is recorded at the end of each step. To update When stronger volume preservation is required, the effect of explicit iterative projection can be achieved by repeatedly and alternately performing "recalculate volume - apply constraint force" in the outer layer iteration.

[0036] (4-3) Finally, calculate the resultant external force based on the above external forces. And calculate the acceleration of the particle. Then, the velocity and position are updated using the Symplectic Euler method to ensure synchronization between control commands and physical responses. in, It is the time step for solving.

[0037] (4-4) This module provides a closed-loop parameter optimization and performance evaluation interface: During system operation, by dynamically adjusting hardware structural parameters (such as elastic stiffness and mass distribution) and control parameters (such as drive gain and tension coefficient), the following dimensions of performance evaluation are achieved: Motion stability evaluation: monitoring whether the system experiences physical oscillation or instability under different operating conditions; Energy consumption evaluation: analyzing the power consumption distribution under different control strategies and optimizing the energy utilization rate of the actuator; Entity consistency verification: comparing the monitoring data with the feedback from the entity sensors to ensure a high-precision fit between the simulation drive logic and the real physical behavior.

[0038] Thus, the construction of a neural-mechanical coupled software multimodal motion control method based on a pre-trained language model is completed.

[0039] Another objective of this invention is to provide the method for motion gait planning and intelligent control of biomimetic soft robots in unstructured environments.

[0040] The technical solution provided by the invention may include the following beneficial effects: (1) Highly Stable Bionic Physical Actuator: Inspired by the multimodal motion behavior of fruit fly larvae, this invention constructs a soft robot body based on deep coupling of "skeleton nodes - elastic constraints - actuator array", realizing high-degree-of-freedom physical deformation of the continuum structure. By adopting a regular hexagonal segment layout and multi-directional elastic support design, the actuator exhibits excellent topological stability while possessing flexible adaptability, effectively solving the technical pain point of traditional soft structures that are difficult to balance deformation control and structural maintenance under large deformation.

[0041] (2) A semantically driven intelligent execution path was realized: This invention constructs a cross-modal "semantic-rhythm-execution" end-to-end automatic mapping mechanism, which integrates pre-trained language models with neurodynamic control and artificial muscle drive systems. By directly converting abstract natural language instructions into physical drive parameters of the underlying actuators, smooth switching and autonomous scheduling of multiple motion modalities such as peristalsis, turning, and rolling are realized under a unified control architecture, which significantly improves the autonomous decision-making ability and environmental adaptation level of the software robot system.

[0042] (3) Dynamic optimization platform with integrated digital twin drive: This invention achieves joint optimization of structural physical parameters and control parameters by integrating a state monitoring and dynamic debugging module. This module is based on a dynamic drive engine and supports closed-loop calculation of multiple indicators such as motion stability, energy consumption efficiency and motion trajectory accuracy. It can complete the pre-verification of different drive strategies before the preparation of physical prototypes, providing a highly reliable physical configuration scheme for hardware implementation.

[0043] (4) Possesses broad engineering applications and technical adaptability: The system and method described in this invention can be directly applied to scenarios such as soft robot design verification, biological motion mechanism research, and neuro-mechanical coupled behavior control. As a general soft robot mechanical control method, this invention provides core technical support for the engineering implementation of soft robots in fields such as medical assistance, intelligent detection in confined spaces, and the development of biomimetic intelligent systems.

[0044] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0045] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0046] Figure 1 This is a flowchart illustrating the overall framework of the present invention.

[0047] Figure 2 This is a schematic diagram of the body structure of the multi-segment soft robot provided by the present invention (representing a mass system, a muscle actuation unit, and a spring connection network).

[0048] Figure 3 This is a diagram showing the changes in segmental neural activity signals and the time step difference of signal transmission between segments when the excitation / inhibition coupling coefficient in the rhythmic oscillation control equation of this invention is large.

[0049] Figure 4 This is a diagram showing the changes in segmental neural activity signals and the time step difference of signal transmission between segments when the excitation / inhibition coupling coefficient in the rhythmic oscillation control equation of this invention is small.

[0050] Figure 5 This is a keyframe diagram of a single forward peristaltic cycle of a multi-segment soft robot in a simulated scenario provided by the present invention.

[0051] Figure 6 These are keyframe images of a multi-segment soft robot tumbling laterally in a simulated scenario provided by this invention.

[0052] Figure 7The keyframe (AD) of a single head-turning cycle of a multi-segment soft robot in a simulated scenario provided by this invention, and the robot's posture (E) after multiple turning cycles. Detailed Implementation

[0053] To better understand the technical solution of the present invention, the embodiments of the present invention will be further described below in conjunction with the accompanying drawings and specific examples. Note that the aspects described below in conjunction with the accompanying drawings and specific examples are merely exemplary and should not be construed as limiting the scope of protection of the present invention in any way.

[0054] Example 1 This example was performed on a computer equipped with an Intel 13th Gen Intel(R) Core(TM) i7-13700K CPU and an Intel(R) UHD Graphics 770 GPU. The operating system version was Windows 11 Professional 24H2, the Python version was 3.12.6, and the Taichi version was 1.7.2.

[0055] This invention provides a neural-mechanical coupled software multimodal motion control method based on a pre-trained language model, the process of which is as follows: Figure 1 As shown, it includes the following steps: Step 1: First, the system reads the coordinates of key morphological points from the two-dimensional abdominal contour data of the fruit fly larva. This data records the width information of the larva's cross-section at different body segments. To ensure the dynamic balance and left-right symmetry of the actuator during movement, the two-dimensional coordinates are regularized: the positions of the head and tail points are fixed, and the cross-sectional radius of the intermediate body segments is corrected using a mirror operation. Based on the regularized coordinates, the two-dimensional coordinates are mapped to three-dimensional space to generate a continuous skeleton containing 62 particles: two particles represent the head and tail respectively; the remaining 60 particles are divided into 10 body segments, each composed of a regular hexagonal cross-section. The hexagonal cross-sections are generated using polar coordinates: a radius is given for each body segment. Six vertices are generated at equal angles of 60°, with the following coordinates: in This indicates the position of the body segment along the front-rear axis, with the radius determined by the longitudinal value of the two-dimensional input coordinates. This ensures that the cross-sectional topology maintains a regular hexagonal shape, enhancing numerical stability. The generated three-dimensional coordinates are stored in `d3_coordinates` as the basic input for soft robot modeling. The final soft robot body structure appears as follows: Figure 2As shown (from left to right: mass distribution, muscle actuator division, and spring connection network of the robot body), the system performs differentiated mass allocation on 62 skeleton nodes, assigning mass weights to each segment node through the configuration list cfg.MS.MASS. This non-uniform mass modeling method enables the actuator to more accurately simulate the inertial characteristics of a real biological organism, providing physical support for subsequent precise dynamic execution.

[0056] Step 2: This step aims to deploy a rhythmic signal generator based on neurodynamic logic to provide the underlying driving rhythms required for the soft robot to perform multimodal motion. The system deploys a rhythmic oscillation network consisting of 11 cascaded control nodes. Each control node includes an excitatory processing unit, an inhibitory processing unit, and a sensor feedback interface. Phase-delayed signals are transmitted between adjacent nodes via bidirectional coupling links, simulating the neural conduction characteristics of biological systems. An external periodic trigger pulse signal is set and applied only to the first control node to initiate a peristaltic wave. The pulse period is set to 4.0 s, the duration of a single pulse is 0.4 s, and the initial intensity is 0.5. To protect the physical actuators from sudden electrical signal impacts, the driving signal undergoes smooth boosting and shaping processing in each cycle, ensuring the continuity of the driving load and the mechanical stability of the system. The control unit employs high-precision state iteration logic, updating the driving state variables of each node in real time with a sampling step size (dt) of 0.001 s.

[0057] The following table lists the main parameter settings used in this embodiment: Table 1 Parameter settings for the rhythmic oscillation control equations With these parameters configured, the rhythmic oscillation network can spontaneously generate stable phase delay waves. The generated rhythmic signals provide the artificial muscle actuator with a time-consistent and amplitude-controllable driving input, thereby supporting the realization of subsequent multimodal movements such as peristalsis, head swinging, and rolling. Figure 3 and Figure 4 These represent the changes in segmental neural activity signals and the time step difference in signal transmission between segments in the rhythmic oscillation network under different parameter configurations.

[0058] Step 3: The rhythmic signals generated in Step 2 are mapped to physical tension loads acting on the soft actuator, and the spring force, damping force, adhesive force, and volume constraint force between particles are calculated through physical simulation. Through the combined effect of these forces, the body maintains stable volume and structural consistency during movement, while exhibiting movement characteristics close to those of a real larva. This process not only ensures simulation accuracy but also makes the generated motion trajectory more consistent with biological laws.

[0059] Step 4: During system operation, natural language input is passed in. The instruction parsing interface receives the natural language input (e.g., "move forward and avoid obstacles while turning"). The built-in pre-trained language model semantically decodes the instruction and generates corresponding control tokens. These tokens directly adjust the internal registers of the rhythm control unit (e.g., frequency, phase, and weights) to achieve dynamic modulation of the underlying drive waveform. Based on the decoded control tokens, the execution system can autonomously switch between the following three typical motion modes: Peristaltic mode: The token configures the longitudinal muscle actuators to sequentially perform wave-like activation, generating an axial contraction wave that propels the entire body forward; Head-swinging and turning mode: The token adjusts the drive gain of the actuators on both sides of the body segment, achieving precise directional adjustment through asymmetric oblique contraction; Tumbling mode: The token activates the coordinated drive of the oblique muscle actuators and the circumferential muscle actuators, generating a helical load around the longitudinal axis, driving the body to perform a tumbling action. All of the above motion modes are completed under the same control architecture, without the need for manual parameter switching. The system achieves a natural connection from high-level intentions to low-level physical actions through semantic tokens, fully demonstrating the technical advantages of this invention in intelligent drive and multimodal autonomous execution of soft robots.

[0060] Step 5: In this step, the invention verifies the motion of the soft robot under simulated hard plane conditions. The system provides real-time visualization and dynamic parameter adjustment interfaces during operation, recording the robot's trajectory and speed and observing its stability during motion. In this scenario, three motion modes were tested sequentially: peristalsis, head-swinging, and rolling. The peristalsis mode is shown below. Figure 5 As shown, by adjusting the frequency and phase delay of the rhythm signal, a longitudinal contraction wave propagating sequentially from head to tail is achieved. Simulation results show that this mode can stably propel the robot forward in a straight line on a plane; the head-turning mode is as follows: Figure 7 As shown, while maintaining overall body stability, activating the muscles on one side of the body to contract periodically segment by segment, while the muscles on the other side contract continuously, generates a left-right swaying motion through the phase difference, which can effectively change the robot's forward direction. Figure 7 The final sub-figure E shows the position and orientation of the body; in roll mode, the longitudinal and circumferential muscles are activated in synergistic drive, successfully achieving a body roll around the longitudinal axis, and returning to a stable state after the roll, as shown in the image. Figure 6 As shown.

[0061] The verification results of this embodiment show that the neuro-mechanical integrated method proposed in this invention can realize the intelligent generation and smooth switching of multiple motion modes under the same architecture and semantic framework, and exhibits good stability and controllability, providing a solid technical verification basis for subsequent extension to unstructured complex environments.

[0062] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0063] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A neural-mechanical coupled software multimodal motion control method based on a pre-trained language model, characterized in that, This method includes the following modules: (S1) The soft robot body structure module based on mass-spring is configured with a multi-segment structure, mimicking the real biological structure of fruit fly larvae. Each segment contains three-dimensional mass physical features, distributed muscle actuators, and a spring connection network. The mass system includes head mass, tail mass, and multiple hexagonal topologically arranged segment mass. The muscle actuators include longitudinal, circumferential, oblique, and end-oriented muscle units. The spring connection network includes longitudinal springs, oblique springs, annular cross springs, and end-reinforcing springs. (S2) The rhythmic oscillation signal generation and parameter control module based on pre-trained language model modulation establishes a rhythmic oscillation network with excitatory-inhibitory neuron coupling. Each oscillation node corresponds to a segment of the robot. It receives external natural language commands, parses and generates control tokens through the pre-trained language model, and then adjusts the parameters of the CPG rhythmic oscillator in real time to generate voltage pulse / rhythmic drive signals with segment phase delay. (S3) The muscle drive and mechanical calculation module based on neuro-mechanical mapping converts rhythmic oscillation signals into activation instructions for different types of muscle execution units, and calculates the active tension of the muscle in combination with the mechanical behavior characteristics of the muscle, as well as various physical actions such as spring force, damping force, friction force, adhesion force, collision force and volume constraint force, to complete the dynamic solution of the particle system. (S4) The simulation and debugging module based on physical simulation calls the physics engine to input the activation state of the muscle execution unit into the dynamics solver. The above physical constraints are used to maintain the stability of the body deformation and control the soft robot to complete the target motion mode under the specified working conditions.

2. The method according to claim 1, characterized in that, In module (S1), the physical skeleton of the soft robot body adopts a regular hexagonal topological layout, arranged step by step with the axial coordinate of the body segments as the reference, and the side length of the hexagon is set by the cross-sectional radius to define the physical boundary, thereby ensuring that the body segments maintain stable mechanical topological constraints under different deformation modes such as longitudinal compression, circumferential contraction and oblique bending. The hexagonal mass points are connected by springs to limit the relative displacement of local mass points and allow continuous and smooth morphological evolution under large deformation.

3. The method according to claim 1, characterized in that, In module (S1), the soft robot includes an actuator array composed of multiple artificial muscles. A unique addressing coding protocol is used to map muscle types to spatial orientations: first, the muscles are classified into four categories based on their structural orientation: longitudinal, circumferential, oblique, and end-oriented; second, a unique number is determined based on their spatial orientation in the cross-sectional or segmental axial coordinate system; finally, a global index is formed by combining the segment number to which the muscle belongs. Through the coding protocol, the control system can independently drive specific actuators rhythmically.

4. The method according to claim 1, characterized in that, In module (S2), the rhythmic oscillation signal generation module includes a rhythmic signal generator that uses neurodynamic control logic to generate rhythmic driving waveforms: in: i Number the body segments. The first i Segmental excitation, inhibition, and sensory feedback variables; a,b,c,d,e,f, The coupling coefficients control the excitation-inhibition connection strength between the current segment and the adjacent segment, the degree of influence of sensory feedback, and the relative weight of signal transmission between adjacent segments, respectively. This is a gain parameter used to adjust the maximum activity level of the excitation and inhibition units; It is a time constant, used to control the response speed of each unit; It is an external pulse input that is triggered periodically only in the head and body segment to initiate the forward propagation of rhythmic waveforms; The nonlinear transfer function of the unit is used to limit the output range of the node and maintain periodicity; The generator outputs a voltage / drive signal with phase delay according to the segment number, and performs real-time motion phase compensation through sensory feedback variables. The above control logic makes the excitation signals of adjacent segments maintain a fixed delay in the time domain, forming a traveling wave pattern from back to front throughout the body. The periodicity comes from the periodic pulse triggering of the first segment and the limit loop characteristics of the nonlinear dynamic equation. The two work together to achieve continuous and stable rhythm generation.

5. The method according to claim 1, characterized in that, The generation and conversion of multiple motion modes in module (S2) is based on the generation and adjustment of rhythmic oscillation control parameters of a pre-trained language model. The parameter modulation process is as follows: the semantic instruction parsing unit in the control system receives natural language instructions from external input, extracts and decodes the semantic features of the instructions using the pre-trained language model, and generates a set of control tokens corresponding to specific motion modes, including rhythmic feature parameters, phase delay coefficients and driving amplitudes, which are used to configure the underlying parameters of the rhythmic signal generator in real time, thereby driving the soft robot to switch gait and change modes in different working conditions. In the peristaltic forward movement mode, the system sequentially triggers excitation signals in the muscle groups of each segment, forming a contraction wave that propagates axially from tail to head. The activation values ​​of the longitudinal muscles in each segment have a certain phase delay, which allows the contraction wave to be stably transmitted and propels the entire body forward. In the head-turning mode, the controller coordinates the different activation modes of the muscles on the left and right sides. The longitudinal muscles on the side of the bending direction contract stably, providing the body's steering force, while the other side provides oscillating signals to generate periodic fluctuations, propelling the robot to turn stably in the new direction. In the tumbling mode, the neural oscillation signals are redistributed to the oblique muscle groups throughout the body and coupled with the longitudinal and circumferential muscles to drive the movement. Specifically, the oblique muscles of each segment are activated alternately, forming a spiral contraction wave along the longitudinal axis. At the same time, the longitudinal and circumferential muscles provide structural support and sectional contraction, ensuring the continuity and stability of the tumbling process. This synergistic effect enables the robot to quickly complete the overall flip, achieving an emergency avoidance action similar to the escape behavior of fruit fly larvae.

6. The method according to claim 1, characterized in that, In module (S3), the rhythmic oscillation electrical signal generated by the rhythm signal generator in module (S2) is acquired. According to the preset neuro-mechanical mapping protocol, the electrical signal intensity is converted into the activation drive command of the corresponding artificial muscle execution unit. The drive command is then combined with the length-tension physical characteristic curve and velocity-tension physical characteristic curve of the software robot body to output the physical tension control value acting on each group of muscle units in real time. This is to overcome environmental resistance and generate the power to drive the robot's displacement. . in This represents the maximum contractile force of the muscle unit; The muscle activation value, obtained from neuromechanical mapping, originates from the output of rhythmic oscillation signals. This is a length-tension relationship function used to describe the tension changes of muscles at different elongation lengths; This is a velocity-tension relationship function, reflecting the influence of muscle contraction velocity on mechanical output; The corresponding parallel spring force ( (The spring stiffness coefficient) simulates the passive elasticity of muscles and their associated tissues, ensuring that they can still provide restorative force when not activated; For the damping force term ( (This is the damping coefficient), which reflects the viscous drag effect and is used to suppress excessively rapid length changes and avoid numerical instability. This refers to the amount of muscle elongation. This refers to the contraction speed.

7. The method according to claim 1, characterized in that, In module (S4), the simulation and debugging module provides real-time motion monitoring, recording the motion trajectory and performance indicators of the soft robot under different parameter combinations. Specifically, within each simulation time step, the velocity and position of the mass point are solved and updated based on the resultant force of various external forces, ensuring the continuity and stability of the system state evolution over time. Through this module, the effects of different control parameters can be quickly evaluated in a unified simulation platform, realizing the verification of multiple motion modes and providing a reliable theoretical basis for subsequent hardware design and optimization.

8. The method according to claim 7, characterized in that, The various external forces include one or more of the following: muscle contraction force, spring force, damping force, friction force, adhesion force, volume constraint force, and collision reaction force.

9. The method of claim 1 is used for motion gait planning and intelligent control of biomimetic soft robots in unstructured environments.

Images