Adaptive motion generation method and system for a heterogeneous modular robot

Abstract

The application relates to a self-adaptive motion generation method and system of a heterogeneous modular robot, a configuration graph describing the topological relationship of the heterogeneous modular robot is acquired, and the configuration graph is parsed and converted into a full-order dynamics model; based on the full-order dynamics model, an online model prediction control framework is used to minimize a cost function dynamically generated based on the configuration graph at each control time, and a control action sequence covering a future limited time window is optimized online through a multi-step iterative refining process; a first action in the optimized control action sequence is extracted and executed, and the remaining control action sequence is moved forward in time as an initial nominal control action sequence of an iterative refining process at a next control time, so that continuous rolling horizon control is realized. Compared with the prior art, the application realizes adaptive control of any heterogeneous modular robot in a plug-and-play mode, and significantly improves the flexibility, robustness and practicability of the modular robot system.

Description

Technical Field

[0001] This invention relates to the fields of robotics and adaptive control, and in particular to a method and system for generating form-independent adaptive motion for heterogeneous modular robots. Background Technology

[0002] With the development of modular robotics technology, it has become possible to construct robots with specific functions on demand. However, after physical assembly, the core challenge in achieving higher levels of robot autonomy lies in enabling these newly generated, diverse robots to act autonomously—that is, achieving a seamless transition from body construction to behavioral control. The key to endowing new configurations with intelligent behavior lies in their motion generation and control algorithms.

[0003] Currently, methods for solving control problems of novel robot types can be mainly divided into two categories: reinforcement learning (RL)-based methods and model predictive control (MPC)-based methods. Reinforcement learning methods, through extensive trial-and-error training in simulation environments, can learn control policies applicable to complex, contact-rich tasks. However, these methods typically consume enormous computational resources and involve lengthy training times, and the trained policies often overfit the training environment and task, severely limiting their generalization ability and adaptability when faced with unseen dynamic changes (such as increased load) or new tasks. Model predictive control (MPC) methods, especially gradient-based nonlinear MPC (NMPC), demonstrate good real-time feedback capabilities by solving for the optimal control sequence in the future time domain through online optimization. However, to ensure computational efficiency, these methods often rely on simplified reduced-order models (such as single-rigid-body models or center-of-mass dynamics models), sacrificing accurate modeling of the robot's full-order dynamics. This leads to performance degradation in tasks requiring high agility and may even fail to meet complex contact constraints. Directly performing online gradient optimization on the full-order dynamic model faces the problems of optimizing non-convex and non-smooth landscapes and excessive computational cost, making it difficult to achieve real-time solutions.

[0004] As an alternative, sampling-based MPC methods (such as Model Predictive Path Integral, MPPI) have attracted attention due to their advantages of not requiring gradient information, being easy to parallelize, and being able to handle arbitrary dynamic models. These methods sample by superimposing a large number of random perturbations on the control sequence and update the control policy by weighted averaging based on simulation evaluation results. However, existing sampling-based MPC methods have an inherent core drawback: their performance is highly dependent on a fixed sampling variance. If the sampling variance is too large, although extensive global exploration is possible, it is difficult to converge to the optimal solution, resulting in poor control accuracy and high variance. If the sampling variance is too small, although fine local search is possible, it is prone to getting trapped in local optima and failing to discover better motion strategies, especially in high-dimensional, non-convex optimization problems such as robot motion control, where this contradiction is particularly prominent. Therefore, how to efficiently solve the real-time optimal control problem of robots with arbitrary new configurations, while ensuring global search capability and achieving accurate local convergence, thereby generating stable and high-performance dynamic motion, is a major technical challenge in this field. Summary of the Invention

[0005] The purpose of this invention is to solve the core technical challenge of how to achieve real-time and adaptive motion capabilities for heterogeneous modular robots of arbitrary shapes and dynamic reconfigurability, and to provide an adaptive motion generation method and system for heterogeneous modular robots.

[0006] The objective of this invention can be achieved through the following technical solutions: As a first aspect of the present invention, an adaptive motion generation method for a heterogeneous modular robot is provided, comprising the following steps: Obtain the configuration diagram describing the topological relationships of the heterogeneous modular robot, and analytically convert the configuration diagram into a full-order dynamic model; Based on the full-order dynamics model, an online model predictive control framework is used. At each control moment, the goal is to minimize the cost function dynamically generated based on the configuration diagram. Through a multi-step iterative refinement process, the sequence of control actions covering a finite future time window is optimized online. Extract and execute the first action in the optimized control action sequence, and shift the remaining control action sequence forward in time as the initial nominal control action sequence for the iterative refinement process at the next control moment, thereby achieving continuous rolling time-domain control.

[0007] As a preferred technical solution, the cost function dynamically generated based on the configuration diagram includes cost items related to the combination of module types contained in the configuration diagram: If the configuration diagram contains only the legs formed by active joint modules, then a gait regularity cost term that encourages periodic movement is activated, and the end-foot contact force is rewarded. If the configuration diagram includes a wheel module, then the linear velocity tracking cost term is activated, while the reward for the foot-on-the-air phase is suppressed or removed, and a penalty for wheel slippage is added. If the configuration diagram is a hybrid wheel-leg form, then the gait regularity cost term and the linear velocity tracking cost term are activated simultaneously, and their weights are adjusted through preset rules or learning mechanisms to encourage coordinated movement.

[0008] As a preferred technical solution, the online model predictive control framework employs a multi-step iterative refinement process using a model predictive path integral control algorithm, the specific process of which is as follows: Calculate the sampled covariance moments for actions at future time steps in the current iteration; The candidate control action sequence base is obtained by mixing and sampling the nominal sequence and the historical periodic template constructed based on the historical action buffer; wherein, the historical periodic template is constructed as follows: a segment of a set length is truncated from the historical action buffer and repeated periodically to construct a template sequence; the template sequence is mixed with the current nominal sequence by phase gradient, and then low-pass smoothing noise is superimposed to generate the historical periodic template. Based on the dynamically changing sampling covariance matrix, multiple high-dimensional noise sequences are generated in parallel from the Gaussian distribution, and the noise sequences are superimposed on the candidate action sequence basis obtained by mixed sampling to obtain the candidate action sequence. Forward simulation is performed on all candidate control action sequences, and the cumulative cost of each candidate control action sequence is calculated based on the cost function dynamically generated based on the configuration diagram. The normalized weights are calculated for each noise sequence based on the cumulative cost, and the update increment is obtained by weighted summation of all noise sequences. The update increment is applied to the current nominal control action sequence to obtain the updated nominal control action sequence.

[0009] As a preferred technical solution, a larger sampling range is used in the initial stage of the iterative refining process, and the sampling range is gradually reduced in the subsequent stages of the iterative process.

[0010] As a preferred technical solution, the online model predictive control framework assigns a relatively larger sampling variance to actions that are later in time within a single control action sequence.

[0011] As a preferred technical solution, the method continuously monitors whether the input configuration diagram changes; During stable operation, the initial nominal control action sequence for the next moment is derived from the optimized control action sequence of the current moment. When a change in the configuration diagram is detected, the old nominal control action sequence is immediately discarded and reset to a zero sequence; in the next set number of control cycles, the trajectory-level annealing hyperparameter is temporarily increased. The value of makes the sampling variance of the initial iteration larger.

[0012] As a second aspect of the present invention, an adaptive motion generation system for a heterogeneous modular robot is provided, the system executing the adaptive motion generation method for a heterogeneous modular robot as described above, including: The configuration conversion module is configured to receive a configuration diagram describing the robot's topology and parse and convert the configuration diagram into a full-order dynamic model containing complete links, joints, mass, inertia tensors, and contact properties for use by the parallel physics simulation engine. The dynamic cost function generation module dynamically generates cost functions based on the cost items related to the combination of module types contained in the configuration diagram parsed by the configuration transformation module. The motion generation module, connected to the configuration transformation module, is configured to employ a sampling-based online model predictive control framework. Based on a full-order dynamic model, it optimizes the control action sequence online through a multi-step iterative refinement process, aiming to minimize the cost function dynamically generated from the configuration diagram. The parallel simulation and execution module, connected to the motion generation module, is configured to perform cost evaluation of candidate control action sequences in parallel on the graphics processor, output the first action optimized by the motion generation module to the robot's underlying actuator, and manage the update logic in the rolling time domain.

[0013] As a preferred technical solution, the dynamic cost function generation module is based on cost items related to the combination of module types included in the configuration diagram: If the configuration diagram contains only the legs formed by active joint modules, then a gait regularity cost term that encourages periodic movement is activated, and the end-foot contact force is rewarded. If the configuration diagram includes a wheel module, then the linear velocity tracking cost term is activated, while the reward for the foot-on-the-air phase is suppressed or removed, and a penalty for wheel slippage is added. If the configuration diagram is a hybrid wheel-leg form, then the gait regularity cost term and the linear velocity tracking cost term are activated simultaneously, and their weights are adjusted through preset rules or learning mechanisms to encourage coordinated movement.

[0014] As a preferred technical solution, the multi-step iterative refinement process of the motion generation module adopts a strategy of dynamically adjusting the sampling parameters: Based on the current stage of iterative refinement in the overall iterative steps, adjust the overall scale of the sampling parameters applied to the entire control action sequence. In the early stage of the iteration, use a larger sampling range to enhance the global exploration of the solution space, and in the later stage of the iteration, reduce the sampling range to promote accurate convergence to the optimal solution. Within a single sequence of control actions, a set of sampled parameters, distinct from the upcoming near-term actions, is assigned to plan future actions that are far removed from the current time in the time domain, in order to encourage exploratory planning of long-term behaviors.

[0015] As a preferred technical solution, the parallel simulation and execution module continuously monitors whether the input configuration diagram changes; During stable operation, the initial nominal control action sequence for the next moment is derived from the optimized control action sequence of the current moment. When a change in the configuration diagram is detected, the old nominal control action sequence is immediately discarded and reset to a zero sequence; in the next set number of control cycles, the trajectory-level annealing hyperparameter is temporarily increased. The value of makes the sampling variance of the initial iteration larger.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1) This invention empowers modular robot platforms with unprecedented flexibility, robustness, and practicality by constructing an end-to-end automated process from configuration diagram to adaptive behavior, achieving truly plug-and-play, form-independent generalized control. The proposed configuration transformation module is key to realizing form-independent control. It receives an abstract configuration diagram (Graph) that only describes the robot's topological relationships and parses it in real time to generate a full-order dynamic model that can be immediately used for simulation. This means that no matter what unprecedented form the robot is assembled in (e.g., quadruped, hexapod, wheel-leg hybrid, etc.), the system can understand its physical nature and establish accurate dynamic equations within milliseconds, without relying on any pre-set model library or manual modeling process. This fundamentally realizes a high-dimensional generalization capability for the robot's body structure and topological relationships.

[0017] 2) This invention achieves dynamic adaptation and intelligent generation of task objectives. The proposed dynamic cost function generation module enables the robot's behavior to intelligently adapt to its physical form. Based on the configuration diagram parsed by the configuration transformation module, the cost function used for motion optimization is automatically and intelligently combined and adjusted. For example, when a wheeled module is detected in the configuration, the system automatically activates and increases the weights of rewards related to linear velocity tracking and maintaining wheel-ground contact. This allows the robot's behavior to automatically adapt to its new physical functions, rather than simply moving, thus exhibiting stronger task orientation.

[0018] 3) To address the issues of poor rhythm and excessive randomness in existing MPPI algorithms when handling periodic movements, this invention introduces a unique hybrid distribution sampling strategy during the optimization process. This is a key mathematical improvement for achieving more rhythmic movement, rather than relying solely on cost function constraints. A hybrid distribution containing a historical action buffer is constructed, concentrating the sampling probability quality near the historical gait phase / periodic pattern. Compared to the traditional method that forces the robot to produce periodic movements solely through cost function penalties, this method converges more easily to a solution that maintains the gait under a limited sampling budget, thus exhibiting significantly enhanced rhythm and naturalness.

[0019] 4) This invention achieves robust and rapid adaptation to the dynamic reconfiguration process. The designed reconfiguration adaptive initialization and parameter adjustment mechanism ensures that the robot can quickly recover high-performance motion when its shape changes. When the system detects a change in the robot's topology, it actively triggers two key actions: 1) resetting the control sequence, immediately discarding the local optimum experience learned based on the old shape; 2) temporarily increasing the exploration parameters, forcing the controller to conduct a broader global exploration on the unknown dynamic model of the new shape. This mechanism ensures that the robot can quickly discover and converge to an efficient motion strategy suitable for the new body after reconfiguration, rather than falling into old, no-longer-applicable behavioral patterns.

[0020] 5) This invention achieves efficient online motion planning and generation based on full-order models. A sampling optimization strategy based on diffusion annealing is employed to construct an efficient online motion generation process. This strategy can efficiently solve complex, high-dimensional, non-convex full-order dynamics model optimization problems by dynamically balancing global exploration and local convergence without requiring gradient information. Combined with GPU parallel acceleration, real-time online optimization of complex models becomes possible, enabling the continuous generation of highly agile and high-performance motion commands for robots of arbitrary configurations, completing an end-to-end automated closed loop from abstract configuration to specific adaptive behavior. Attached Figure Description

[0021] Figure 1 This is a flowchart of the morphology-independent adaptive motion generation method for heterogeneous modular robots according to the present invention.

[0022] Figure 2 This is a schematic diagram of a quadruped robot walking in an embodiment of the present invention. Detailed Implementation

[0023] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0024] Example 1 This invention proposes a form-independent adaptive motion generation method and system for heterogeneous modular robots. The core idea of ​​this method is to construct a complete automated process from configuration analysis to real-time control. This process can generate high-quality dynamic motion online and in real-time for any newly assembled robot configuration. The technical solution of this invention is implemented through an integrated system framework. First, the system receives a configuration diagram describing the topology of the heterogeneous modular robot and automatically analyzes it, converting it into a full-order dynamic model that can be used for parallel physical simulation. This model accurately describes all the links, joints, mass, inertia, and contact properties of the robot, providing a solid foundation for subsequent high-fidelity online optimization.

[0025] like Figure 1 As shown in the figure, the form-independent adaptive motion generation method for heterogeneous modular robots according to an embodiment of the present invention includes the following steps in its complete process: Step 1: Automatic configuration conversion and establishment of full-order dynamic model.

[0026] Given a newly assembled or reconstructed heterogeneous modular robot, its physical structure is represented by a configuration diagram describing its topological relationships. Defined as such, this configuration diagram details the type (e.g., basic structure module, active joint module, wheeled module), three-dimensional spatial position, and physical connections between each module that constitutes the robot. The system first automatically parses this configuration diagram through a configuration conversion module. This module transforms abstract graph structure information into a format suitable for parallel physics simulation engines (such as...). or Full-order dynamic model The model This precisely includes all the robot's physical properties, such as the mass, inertia tensor, and collider geometry of each module as a link, as well as the type of joint (e.g., rotary joint, fixed connection), range of motion, and drive limitations) at the connections between modules. This step provides a solid physical foundation for subsequent high-fidelity online optimization, avoiding performance losses caused by model simplification in traditional methods.

[0027] Step 2: Enter the rolling time domain control loop and generate the dynamic cost function.

[0028] This invention employs a sampling-based online model predictive control (MPC) framework, transforming the motion generation problem into a finite-time optimal control problem. At each control time... t (For example, at a frequency of 50 Hz), the goal of the line model predictive control (MPC) framework is to solve for a function that minimizes a predefined cost function. J Optimal control action sequence This cost function aims to minimize the cumulative cost within a finite future time window: in, It is the future Step-by-step control action sequence; Is the future number h The state of the step; Is the future number h Step control actions; and These are stage costs and end-user costs; Its specific composition is dynamic and changes.

[0029] A key technical aspect of this invention is that the cost function is not fixed, but rather generated by a dynamic cost function generation module based on the configuration diagram parsed in step 1. Dynamically generated and weighted. This module checks the module types included in the configuration and automatically combines the relevant cost items.

[0030] For example, if the configuration diagram only contains active joint modules, which is a pure leg configuration, the module will automatically activate a gait regularity cost item that encourages periodic movements. And reward the contact force at the end of the foot. .

[0031] If a wheel module is detected in the configuration diagram, the module will automatically activate a high-weight linear velocity tracking cost term. At the same time, it suppresses or removes the reward for the foot take-off phase, and instead punishes wheel slippage. .

[0032] If the configuration is a hybrid wheel-leg configuration, the module will activate both types of cost items simultaneously and introduce a weighting adjustment factor. The weight balance can be adjusted through preset rules or simple learning mechanisms. and To encourage coordinated movement.

[0033] This dynamic adaptability of the cost function is key to solving the functional diversity of heterogeneous robots, and it is something that existing control methods designed for fixed-form robots do not possess.

[0034] Step 3: Iterative refining and adaptive reconstruction based on diffusion annealing.

[0035] To efficiently solve the aforementioned nonconvex optimization problem, a multi-step iterative refinement process is introduced at each control time step, employing a strategy of dynamically adjusting the sampling parameters. This strategy is inspired by recent research combining control theory with diffusion models. Its core idea is to use a larger sampling range in the early stages of iteration to promote global exploration, while gradually reducing the sampling range in the later stages to promote local convergence.

[0036] Specifically, for the nominal control action sequence implement N Sub-iteration refinement (e.g.) ), in the i In the next iteration ( i From 1 to N ), perform the following sub-steps: 3.a) Calculate the annealing sampling covariance: in the first... In the next iteration, for the future... Step action calculation sampling covariance matrix At that time, the following double-cycle diffusion annealing mechanism is adopted: in, and These are hyperparameters that control the trajectory-level annealing and action-level annealing rates, respectively. It is the identity matrix. This mechanism ensures that the sampling variance decreases with the number of iterations. The increase in variance leads to a systematic decrease, achieving trajectory-level refinement. In the early stages of iteration, the sampling variance is relatively large for global exploration; in the later stages of iteration, the sampling variance decreases for local refinement. Simultaneously, within a single control action sequence, actions that occur later in time are considered later time steps. Assign a relatively larger sampling variance to encourage long-term exploration.

[0037] 3.b) Parallel Sampling and Cost Assessment: To address the issues of poor rhythm and excessive randomness in existing Predictive Path Integral (MPPI) algorithms when handling periodic motion, this invention introduces a unique hybrid distribution sampling strategy during the optimization process. This is a key mathematical improvement for achieving more rhythmic motion, rather than relying solely on cost function constraints. Instead of using a single Gaussian distribution, this invention constructs the following hybrid distribution, including a historical action buffer, to sample candidate sequences: in, Indicates a mixed distribution; Indicates the relationship with the current nominal sequence Sampling is performed as a basis for candidate action sequences (stable and easy to converge). This indicates that the candidate action sequence is sampled using a periodic template constructed from historical actions. Specifically, it is derived from the historical action buffer. Constructing a periodic template and the current nominal sequence Phase gradient blending is performed to obtain Then, low-pass smoothing noise is superimposed to generate historical periodic templates (providing rhythmic priors and improving periodicity and naturalness). This is the mixing coefficient, which is set to 0.5 in this embodiment.

[0038] This invention addresses the rhythmicity of robots by using a periodic template constructed from historical actions as a basis to generate candidate action sequence bases through noise addition. Specific improvements are as follows: Periodic template construction: utilizing historical action buffers The cut length is The fragments are periodically repeated, and the construction length is... template sequence : in, Represents the historical action buffer The current length or end index (used to truncate a length of [value] from the end of the buffer). (segment) For modulo operation; is the length of the template sequence.

[0039] This was introduced directly at the sampling source. The physical periodicity constraint ensures that the generated template has strict periodicity, providing a sense of rhythm for subsequent actions.

[0040] 2. Phase gradient blending: To avoid blending historical actions with the current nominal sequence. Phase breakage occurs; a linear gradient coefficient is introduced. : in, This is a sequence after phase gradient mixing; The mixing coefficients are linearly gradually varying, from Gradually increase to 1.

[0041] 3. Low-frequency disturbance injection: Unlike traditional white noise, this scheme superimposes smoothed noise after FIR low-pass filtering to preserve low-frequency parameter tuning characteristics and suppress high-frequency jitter. in, As the basis for the injected perturbation of candidate action sequences; is the random amplitude factor of the noise; This is the time-varying noise scaling factor; This is a smoothed noise sequence after FIR low-pass filtering.

[0042] The method proposed in this invention focuses the sampling probability quality on the vicinity of historical gait phase / periodic patterns. Compared to traditional methods that rely solely on cost function penalties to force robots to produce periodic movements, this method converges more easily to a gait-maintaining solution under a finite sampling budget, thus exhibiting significantly enhanced rhythm and naturalness.

[0043] After obtaining the basis of candidate action sequences through hybrid sampling, multiple high-dimensional noise sequences are generated in parallel from a Gaussian distribution based on the dynamically changing covariance matrix. And these noise sequences are superimposed on the mixed sampling result. K Candidate action sequences are obtained from a base of candidate action sequences.

[0044] Subsequently, using a parallel simulation environment accelerated by a graphics processing unit (GPU), this... K Simultaneously perform forward simulation on each candidate control action sequence, and calculate the cumulative cost of each candidate control action sequence based on the cost function dynamically generated in step 2. .

[0045] 3.c) Update the nominal control action sequence: Based on the update rule of the Model Predicted Path Integral (MPPI), calculate the normalized weights for each noise sequence based on the cost value. : in, It's a temperature parameter. Indicates the first k The cumulative cost of each candidate control action sequence.

[0046] By weighted summation of all noise sequences This yields an update increment, which is then applied to the current nominal control action sequence to obtain the updated nominal control action sequence after this iteration. .

[0047] Another key technology of this invention lies in the reconstruction of the adaptive initialization and parameter adjustment mechanism. The system continuously monitors the input configuration diagram. Has anything changed?

[0048] During stable operation, the initial nominal control action sequence for the next moment is derived from the optimized control action sequence of the current moment.

[0049] When a change in the configuration diagram is detected (i.e., the robot is reconfigured), the system determines that the dynamic model has been fundamentally altered. At this point, the system will: Reset: Immediately discard the old nominal control action sequence and reset it to a zero sequence.

[0050] Enhanced exploration: Temporarily increase the trajectory-level annealing hyperparameters during the next few control cycles. The value of this will make the sampling variance of the initial iteration larger, forcing the controller to explore new and unknown dynamic models more extensively, thereby quickly adapting to new forms and avoiding getting trapped in local optima based on old forms.

[0051] Step 4: Action execution and time-domain scrolling.

[0052] After completion N After several iterations of refinement, the final optimized control action sequence is obtained. The first action of this sequence. The control is sent to the robot's underlying joint torque controller for execution. Then, the entire control sequence is rolled forward one step to become the next control moment. t The initial value is refined by iteratively adding 1 to achieve continuous and smooth robot motion.

[0053] Example 2 As one embodiment of the present invention, the present invention also proposes a form-independent adaptive motion generation system for heterogeneous modular robots that performs the above method embodiments, the system comprising: The configuration conversion module is configured to receive a configuration diagram describing the robot's topology and automatically parse and convert it into a full-order dynamic model containing complete links, joints, mass, inertia tensors, and contact properties for use by the parallel physics simulation engine.

[0054] The dynamic cost function generation module dynamically generates cost functions based on the cost items related to the combination of module types contained in the configuration diagram parsed by the configuration transformation module.

[0055] A motion generation module, connected to the configuration transformation module, is configured to employ a sampling-based online model predictive control framework and, based on the full-order dynamics model, optimize a control action sequence online through a multi-step iterative refinement process. This refinement process employs a strategy of dynamically adjusting sampling parameters. This strategy systematically adjusts the overall scale of the sampling parameters applied to the entire control action sequence according to the stage of the current iteration in the overall iteration steps. A larger sampling range is used in the early stages of iteration to enhance global exploration of the solution space, while the sampling range is reduced in the later stages of iteration to promote accurate convergence to the optimal solution. Within a single control action sequence, a set of sampling parameters, distinct from the upcoming near-term actions, is assigned to plan future actions that are far removed from the current time in the time domain, encouraging exploratory planning of long-term behaviors.

[0056] The parallel simulation and execution module, connected to the motion generation module, is configured to perform cost evaluation of a large number of candidate control action sequences in parallel on a graphics processing unit (GPU), output the first action optimized by the motion generation module to the robot's underlying actuator, and manage the update logic in the rolling time domain.

[0057] This embodiment demonstrates a dynamic reconfiguration and adaptive control process from a quadruped robot to a wheeled robot.

[0058] Initial state: The controlled object is a quadruped robot assembled from 12 active joint modules and several basic structural modules. System response: The configuration transformation module generates a dynamic model of the quadruped morphology. The dynamic cost function module detects that only joint modules are present, therefore activating a combination of cost terms used to maintain body balance, track target velocity, and achieve efficient trotting gait to enable the robot to walk stably, such as... Figure 2 As shown. Reconfiguration process: An external assembly robot adds two wheel modules to the ends of the front and rear legs of the quadruped robot, respectively. Post-reconfiguration state: The robot's form changes to a hybrid wheel-leg configuration. The system receives a new configuration diagram.

[0059] Real-time model update: The configuration conversion module automatically generates a new full-order dynamic model of the wheeled robot within one control cycle (<20ms).

[0060] Adaptive cost function: The dynamic cost function module detects the newly added "wheel module" and immediately adjusts the cost function: it reduces the weight of leg gait regularity, while activating and assigning high weights to the cost terms of "precise linear velocity tracking" and "maintaining wheel contact with the ground".

[0061] Reconfiguration Adaptive Exploration: The controller detects a significant change in the configuration diagram and triggers the reconfiguration adaptive mechanism. It resets the nominal control action sequence to zero and temporarily increases the annealing parameters. In the subsequent iterative refinement, the first iteration ( i =1) The sampling range became very large, and the controller explored a variety of possibilities, including: "walking on legs while the wheels are spinning", "rolling on wheels while the legs are stiff", and "coordinated wheel and leg drive".

[0062] Fast convergence: Because the cost function explicitly rewards rolling behavior, after... N Through four iterations, from rough exploration to fine optimization, the controller quickly converged to an efficient motion strategy of "primarily using wheel drive, with leg joints assisting in posture adjustment and obstacle crossing".

[0063] Final results: Within seconds of the reconstruction being completed, the robot seamlessly switched from a walking gait to a rolling posture and began precisely tracking the new speed commands. The entire process was fully automated, requiring no human intervention or offline retraining.

[0064] The above description represents the preferred embodiments of this application. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this application, and these improvements and modifications are also considered within the scope of protection of this application. For example, the cost function can be adjusted according to the specific task, and the underlying controller can be adapted to position control or hybrid control modes.

[0065] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0066] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. An adaptive motion generation method for a heterogeneous modular robot, characterized by the following steps: include: Obtain the configuration diagram describing the topological relationships of the heterogeneous modular robot, and analytically convert the configuration diagram into a full-order dynamic model; Based on the full-order dynamics model, an online model predictive control framework is used. At each control moment, the goal is to minimize the cost function dynamically generated based on the configuration diagram. Through a multi-step iterative refinement process, the sequence of control actions covering a finite future time window is optimized online. Extract and execute the first action in the optimized control action sequence, and shift the remaining control action sequence forward in time as the initial nominal control action sequence for the iterative refinement process at the next control moment, thereby achieving continuous rolling time-domain control.

2. The adaptive motion generation method for a heterogeneous modular robot according to claim 1, characterized in that, The cost function dynamically generated based on the configuration diagram includes cost items related to the combination of module types contained in the configuration diagram: If the configuration diagram only contains the legs formed by active joint modules, then the gait regularity cost term that encourages periodic movement is activated, and the end-foot contact force is rewarded. If the configuration diagram contains only wheel modules, then the linear velocity tracking cost term is activated, while the reward for foot take-off is suppressed or removed, and a penalty for wheel slippage is added. If the configuration diagram is a hybrid wheel-leg configuration, then the gait regularity cost term and the linear velocity tracking cost term are activated simultaneously, and the weight adjustment factor adjusts the weights of the two according to the increase or decrease of the active joint module and the wheel module to encourage coordinated movement.

3. The adaptive motion generation method for a heterogeneous modular robot according to claim 1, characterized in that, The online model predictive control framework employs a multi-step iterative refinement process using the model predictive path integral control algorithm, as detailed below: Calculate the sampled covariance moments for actions at future time steps in the current iteration; The candidate control action sequence base is obtained by mixing and sampling the nominal sequence and the historical periodic template constructed based on the historical action buffer; wherein, the historical periodic template is constructed as follows: a segment of a set length is truncated from the historical action buffer and repeated periodically to construct a template sequence; the template sequence is mixed with the current nominal sequence by phase gradient, and then low-pass smoothing noise is superimposed to generate the historical periodic template. Based on the dynamically changing sampling covariance matrix, multiple high-dimensional noise sequences are generated in parallel from the Gaussian distribution, and the noise sequences are superimposed on the candidate action sequence basis obtained by mixed sampling to obtain the candidate action sequence. Forward simulation is performed on all candidate control action sequences, and the cumulative cost of each candidate control action sequence is calculated based on the cost function dynamically generated based on the configuration diagram. The normalized weights are calculated for each noise sequence based on the cumulative cost, and the update increment is obtained by weighted summation of all noise sequences. The update increment is applied to the current nominal control action sequence to obtain the updated nominal control action sequence.

4. The adaptive motion generation method for a heterogeneous modular robot according to claim 3, characterized in that, A larger sampling range is used in the initial stage of the iterative refining process, and the sampling range is gradually reduced in the subsequent stages of the iterative process.

5. The adaptive motion generation method for a heterogeneous modular robot according to claim 3, characterized in that, The online model predictive control framework assigns a relatively larger sampling variance to actions that are later in time within a single control action sequence.

6. The adaptive motion generation method for a heterogeneous modular robot according to claim 1, characterized in that, The method continuously monitors whether the input configuration diagram changes; During stable operation, the initial nominal control action sequence for the next moment is derived from the optimized control action sequence of the current moment. When a change in the configuration diagram is detected, the old nominal control action sequence is immediately discarded and reset to a zero sequence; in the next set number of control cycles, the trajectory-level annealing hyperparameter is temporarily increased. The value of makes the sampling variance of the initial iteration larger.

7. An adaptive motion generation system for a heterogeneous modular robot, characterized in that, The system executes the adaptive motion generation method for heterogeneous modular robots as described in any one of claims 1-6, including: The configuration conversion module is configured to receive a configuration diagram describing the robot's topology and parse and convert the configuration diagram into a full-order dynamic model containing complete links, joints, mass, inertia tensors, and contact properties for use by the parallel physics simulation engine. The dynamic cost function generation module dynamically generates cost functions based on the cost items related to the combination of module types contained in the configuration diagram parsed by the configuration transformation module. The motion generation module, connected to the configuration transformation module, is configured to employ a sampling-based online model predictive control framework. Based on a full-order dynamic model, it optimizes the control action sequence online through a multi-step iterative refinement process, aiming to minimize the cost function dynamically generated from the configuration diagram. The parallel simulation and execution module, connected to the motion generation module, is configured to perform cost evaluation of candidate control action sequences in parallel on the graphics processor, output the first action optimized by the motion generation module to the robot's underlying actuator, and manage the update logic in the rolling time domain.

8. The adaptive motion generation system for a heterogeneous modular robot according to claim 7, characterized in that, The dynamic cost function generation module is based on cost items related to the combination of module types included in the configuration diagram: If the configuration diagram contains only the legs formed by active joint modules, then a gait regularity cost term that encourages periodic movement is activated, and the end-foot contact force is rewarded. If the configuration diagram includes a wheel module, then the linear velocity tracking cost term is activated, while the reward for the foot-on-the-air phase is suppressed or removed, and a penalty for wheel slippage is added. If the configuration diagram is a hybrid wheel-leg form, then the gait regularity cost term and the linear velocity tracking cost term are activated simultaneously, and their weights are adjusted through preset rules or learning mechanisms to encourage coordinated movement.

9. The adaptive motion generation system for a heterogeneous modular robot according to claim 7, characterized in that, The multi-step iterative refinement process of the motion generation module employs a strategy of dynamically adjusting sampling parameters. Based on the current stage of iterative refinement in the overall iterative steps, adjust the overall scale of the sampling parameters applied to the entire control action sequence. In the early stage of the iteration, use a larger sampling range to enhance the global exploration of the solution space, and in the later stage of the iteration, reduce the sampling range to promote accurate convergence to the optimal solution. Within a single sequence of control actions, a set of sampled parameters, distinct from the upcoming near-term actions, is assigned to plan future actions that are far removed from the current time in the time domain, in order to encourage exploratory planning of long-term behaviors.

10. The adaptive motion generation system for a heterogeneous modular robot according to claim 7, characterized in that, The parallel simulation and execution module continuously monitors whether the input configuration diagram changes; During stable operation, the initial nominal control action sequence for the next moment is derived from the optimized control action sequence of the current moment. When a change in the configuration diagram is detected, the old nominal control action sequence is immediately discarded and reset to a zero sequence; in the next set number of control cycles, the trajectory-level annealing hyperparameter is temporarily increased. The value of makes the sampling variance of the initial iteration larger.

Images