Multi-form robot motion control system based on AI perception decision

Abstract

The invention discloses a multi-form robot motion control system based on AI perception decision. The multi-form robot motion control system comprises a perception fusion module, a multi-mode prediction module, a form management module, a motion control module and a control optimization module. The invention relates to the technical field of robot intelligent control, in particular to a polymorphic robot motion control system based on AI perception decision-making, according to the scheme, discrete morphologies are linearly combined into unified hybrid dynamics through micro-hybrid weights, and a multi-modal predicted safety violation probability and entropy regularization learning morphological strategy are combined to obtain a multi-modal robot motion control system. Smooth low-frequency switching and continuous balancing of energy consumption and risks are achieved, and coupling with model prediction control is achieved; the multi-modal collision probability is merged into model prediction control cost and constraint, hybrid dynamics is adopted for solving, high-frequency whole-body inverse dynamics is mapped into constrained joint torque, safety filtering is combined, priority avoidance and smooth executable control are guaranteed, energy consumption is reduced, and the task success rate and safety are improved.

Description

Technical Field

[0001] This invention relates to the field of robot intelligent control technology, specifically to a multi-morphic robot motion control system based on AI perception and decision-making. Background Technology

[0002] With the decrease in cost and the improvement in computing performance of LiDAR, depth cameras, inertial measurement units, and other multi-source sensors, robot perception capabilities have evolved from traditional geometric localization and simultaneous localization and mapping to dense semantic perception with uncertainty quantification. Deep learning methods and sequence modeling techniques have made significant progress in target detection, semantic segmentation, and long-term trajectory prediction, enabling multimodal perception results to provide richer contextual information for the decision layer. At the same time, optimization-based motion planning methods, such as model predictive control, and whole-body inverse dynamics and contact force calculation techniques for high-frequency execution are being integrated into an end-to-end perception-to-control pipeline. Control obstacle functions and online adaptive learning techniques provide the theoretical and engineering foundation for ensuring the safety and robustness of robot control.

[0003] Existing research on multimorphic robots typically treats robot morphology as discrete switching and employs threshold-based or rule-based switching strategies. This approach suffers from several problems: abrupt changes in dynamics during switching lead to discontinuities in planning and control; high switching jitter and mechanical wear; difficulty in continuously balancing task completion rate, energy consumption, and predicted safety risks; and difficulty in directly integrating dynamic models of different morphologies. Existing motion control methods often only perform model predictive control under a single most probable trajectory or simple obstacle avoidance cost. This approach cannot systematically handle multimodal prediction uncertainties in dynamic environments; it is difficult to identify and prioritize the avoidance of high-risk targets; and it cannot simultaneously satisfy dynamic and contact constraints at the whole-body execution level, leading to conflicts between safety and tracking performance. Summary of the Invention

[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a multi-morphic robot motion control system based on AI perception and decision-making. Current multi-morphic robot research typically treats robot morphology as discrete switching and employs threshold- or rule-based switching strategies. This leads to problems such as discontinuous planning and control due to abrupt dynamic changes during switching, high switching jitter and mechanical wear, difficulty in continuously balancing task completion rate, energy consumption, and predicted safety risks, and difficulty in directly integrating dynamic models of different morphologies. This solution uses differentiable hybrid weights to linearly combine discrete morphologies into a unified hybrid dynamics. Furthermore, it introduces safety violation probabilities and entropy regularization from multimodal predictions and learns morphology strategies during morphology optimization, making morphology selection a differentiable variable that can be jointly optimized with control. This achieves smooth, low-frequency morphology transitions, reduces switching energy consumption and mechanical shock, and continuously balances task performance and risk constraints. Existing motion control methods typically only perform model predictive control under a single most probable trajectory or simple obstacle avoidance cost. The existing methods suffer from several problems: inability to systematically handle multimodal prediction uncertainties in dynamic environments, difficulty in identifying and prioritizing the avoidance of high-risk targets, and inability to simultaneously satisfy dynamic and contact constraints at the whole-body execution level, leading to conflicts between safety and tracking performance. This solution addresses these issues by weighting the multimodal trajectory probability of each dynamic target with a feature-based learnable risk attention and incorporating it into the overall system. To address the costs and constraints, hybrid dynamics is used to support morphological coupling. Subsequently, high-frequency whole-body inverse dynamics quadratic programming is used to map the desired task acceleration into joint torques that satisfy dynamic and contact constraints, enabling safety filtering and online adaptive parameter updates.

[0005] The technical solution adopted by the present invention is as follows: The present invention provides a multi-morphic robot motion control system based on AI perception and decision-making, including a perception fusion module, a multi-modal prediction module, a morphology management module, a motion control module, and a control optimization module;

[0006] The perception fusion module integrates data from multiple heterogeneous sensors, outputs an uncertain state estimate with covariance and a dense semantic occupancy map, and simultaneously constructs training data.

[0007] The multimodal prediction module is used to perform long-term trajectory prediction of dynamic obstacles in the environment and generate multimodal future distributions.

[0008] The morphology management module performs long-term, multimodal trajectory prediction for dynamic targets in the environment. It uses Gaussian mixture to output modal weights, mean trajectories, and covariance for each target to represent the future uncertainty distribution and to determine the optimal morphology and hybrid dynamic parameters of the robot.

[0009] The motion control module operates risk attention-driven model predictive control under hybrid dynamics. The attention-weighted multimodal collision probability for each target is incorporated into the cost or constraint. The expected task space acceleration is then mapped to joint torque by high-frequency whole-body inverse dynamics quadratic programming. After satisfying the dynamics and contact constraints, whole-body motion control commands are generated.

[0010] The control optimization module performs security filtering of underlying instructions and online adaptive updates of system parameters.

[0011] Furthermore, the perception fusion module, which fuses multi-source heterogeneous data and outputs a state estimate with covariance and a dense semantic occupancy map, specifically includes the following units:

[0012] The multimodal data acquisition unit utilizes the data carried on the robot body. Depth camera The robot uses wheel speed sensors to collect raw data in various terrains and records the robot's actual pose and the trajectory of dynamic obstacles in the environment as supervisory labels.

[0013] The multi-sensor probabilistic fusion unit fuses multi-sensor observations within a Bayesian framework to compute the state posterior; it uses Jacobi linearization for nonlinear observations, outputs the state and covariance, first performs posterior factorization, and then performs extended Kalman filter measurement updates.

[0014] Occupy grid mapping units, accumulate occupation information through log probability, and maintain voxel semantic distribution by log probability accumulation; use sparse octree structure to perform incremental read and write, first update the occupation log probability; then perform semantic log confidence accumulation and normalization.

[0015] Furthermore, the multimodal prediction module specifically includes the following units:

[0016] The multimodal trajectory prediction unit uses a Gaussian mixture model to generate a multimodal future trajectory distribution for each dynamic target, and outputs the modal weights, mean trajectory, and covariance.

[0017] Furthermore, the morphology management module specifically includes the following units:

[0018] Establish Markov decision-making processes for high-level task decision-making units. ,by The form defines high-level task decisions, with the goal of maximizing the cumulative reward of the discount. The action set includes motion objectives and form switching decisions.

[0019] Morphological optimization unit, using continuous mixed weights Discrete morphologies are linearly combined into a differentiable hybrid dynamics model, while the morphology optimization objective incorporates the expected performance of the task, the probability of safety violations obtained from the prediction, and entropy regularization.

[0020] Furthermore, the motion control module specifically includes the following units:

[0021] Model predictive control unit, in finite time domain model predictive control The method introduces attention weights for multiple targets in the environment, and incorporates the attention-weighted collision probabilities of each target's multimodal predictions into the algorithm. Constraints enable real-time motion planning with learnable priorities for high-risk or important targets;

[0022] The whole-body inverse dynamics unit solves a quadratic programming problem at high frequencies, and... The desired task space acceleration is mapped to joint acceleration or torque, and tracking error and execution cost are minimized under equation dynamics and contact constraints. The output torque command can be directly issued to the driver.

[0023] Furthermore, the control optimization module specifically includes the following units:

[0024] The safety filtering unit uses a control barrier function at the front end of the control link to transform the safety set into a linear constraint on the input, and then uses real-time quadratic programming. Make minimal modifications to the reference control to obtain the execution input that satisfies the safety constraints;

[0025] Adaptive learning unit, utilizing The optimizer performs small-step online updates to the perception model, dynamic error model, and policy network.

[0026] The beneficial effects achieved by the present invention using the above solution are as follows:

[0027] (1) Existing multi-mode robot research usually treats robot modes as discrete switching and adopts threshold-based or rule-based switching strategies. This results in problems such as discontinuous planning and control due to dynamic abrupt changes during switching, high switching jitter and mechanical wear, difficulty in making continuous trade-offs between task completion rate, energy consumption and predicted safety risks, and difficulty in directly connecting dynamic models of different modes. This solution uses differentiable hybrid weights to linearly combine discrete modes into a unified hybrid dynamics. In the mode optimization, it introduces the safety violation probability and entropy regularization from multi-modal prediction and learns the mode strategy, making mode selection a differentiable variable that can be jointly optimized with control. This achieves smooth, low-frequency mode transitions, reduces switching energy consumption and mechanical shock, and makes continuous trade-offs between task performance and risk constraints, thereby improving the task success rate and lifespan in multi-mode scenarios. At the same time, it is easy to couple with the lower-level MPC in real time to ensure executability.

[0028] (2) Existing motion control methods typically perform MPC only under a single most probable trajectory or simple obstacle avoidance cost, which has the problems of being unable to systematically handle the uncertainty of multimodal prediction in dynamic environments, difficulty in identifying and prioritizing the avoidance of high-risk targets, and the inability to simultaneously satisfy dynamic and contact constraints at the whole-body execution level, resulting in conflicts between safety and tracking performance. This solution incorporates the multimodal trajectory probability of each dynamic target into the cost and constraints of MPC after weighting it with feature-based learnable risk attention, and adopts hybrid dynamics to support morphological coupling. Subsequently, high-frequency whole-body inverse dynamics quadratic programming is used to map the desired task acceleration into joint torque that satisfies dynamic and contact constraints, and safety filtering and online adaptive parameter updates are performed to achieve priority avoidance, safety assurance and smooth executable control in multi-target, high-uncertainty scenarios. Attached Figure Description

[0029] Figure 1 A schematic diagram of a multi-morphic robot motion control system based on AI perception and decision-making provided by the present invention;

[0030] Figure 2 This is a schematic diagram of the perception fusion module;

[0031] Figure 3 This is a schematic diagram of the shape management module;

[0032] Figure 4 This is a schematic diagram of the motion control module;

[0033] Figure 5 This is a schematic diagram of the control optimization module.

[0034] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0035] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0036] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0037] Example 1, see Figure 1 The present invention provides a multi-morphic robot motion control system based on AI perception and decision-making, including a perception fusion module, a multi-modal prediction module, a morphology management module, a motion control module, and a control optimization module;

[0038] The perception fusion module fuses... Depth camera Wheel speed gauge The multi-source heterogeneous sensor outputs an uncertain state estimate with covariance and a dense semantic occupancy map, while simultaneously constructing training data and sending the data to the multimodal prediction module;

[0039] The multimodal prediction module receives data sent by the perception fusion module, performs long-term trajectory prediction of dynamic obstacles in the environment, generates multimodal future distribution, and sends the data to the morphology management module.

[0040] The morphology management module receives data sent by the multimodal prediction module, performs long-term, multimodal trajectory prediction for dynamic targets in the environment, uses Gaussian mixture to output modal weights, mean trajectories and covariance for each target to represent the future uncertainty distribution, and decides the optimal morphology and hybrid dynamic parameters of the robot based on task requirements and environmental risks, and sends the data to the motion control module.

[0041] The motion control module receives data from the morphology management module and operates under a risk attention-driven mechanism in a hybrid dynamics environment. The attention-weighted multimodal collision probability for each target is incorporated into the cost or constraint. The expected task space acceleration is then mapped to joint torque by high-frequency whole-body inverse dynamics quadratic programming. After satisfying the dynamics and contact constraints, whole-body motion control commands are generated and the data is sent to the control optimization module.

[0042] The control optimization module receives data sent by the motion control module and performs security filtering of underlying commands and online adaptive updates of system parameters.

[0043] Example 2, see Figure 1 and Figure 2This embodiment is based on the above embodiment. The perception fusion module fuses multi-source heterogeneous data and outputs a state estimate with covariance and a dense semantic occupancy map. Specifically, it includes the following units:

[0044] The multimodal data acquisition unit employs continuous conductive shielding for sensitive electronic modules and transmission cables, specifically utilizing a metal enclosure, partial shielding covers, and... Conductive sealing strips shield low-impedance connections to the machine ground; surge and transient protection devices are installed at the power inlet. and common-mode chokes; coaxial differential transmission is used for the signal link, utilizing the robot's onboard... Depth camera And wheel speedometer, collecting raw data in various terrains. ;pass Record the robot's actual pose and environmental dynamic obstacle trajectory As a supervisory label; to pair data Used to train the multimodal prediction module, where For the observation context; The sequence is stored in the experience replay pool for training the reinforcement learning strategy in the morphology management module;

[0045] The multi-sensor probabilistic fusion unit fuses multi-sensor observations within a Bayesian framework to compute the posterior state. For nonlinear observations, Jacobi linearization is applied. The output state and covariance are first factorized posteriorly, as shown below: Then, the extended Kalman filter measurement update is performed, as shown below:

[0046] ;

[0047] in, This represents the robot's state vector, which includes the robot's position. ,speed Posture Quaternions , deviation and , Indicates the transpose symbol. Represents the generalized state update operator; This represents the set of observations from the k-th sensor, including IMU acceleration and angular velocity, camera images, depth, LiDAR point cloud, wheel speed, torque sensor readings, and sensor outputs for GPS positioning and target detection. Each observation includes a timestamp and corresponding covariance. Describe the prior distribution, Indicates that in a given set of observations The posterior distribution of the state. Let the observation likelihood function of the k-th sensor be denoted as , specifically, the state is . When the observed value is obtained The probability density; Indicates to The posterior estimate refers to the estimate that has been updated using the current observations. Indicates to Prior estimates refer to estimates that have not yet been updated with current measurements; and Let these represent the prior and posterior covariance matrices, respectively. This represents the Kalman gain matrix, with size . , The dimension representing the robot's state vector; Indicates the observation model in Jacobian matrix at the location, Represents the observation noise covariance matrix. Represents the observation mapping function, Represents the identity matrix with size . ;

[0048] Occupying grid mapping units involves accumulating occupancy information through log probability and maintaining the voxel semantic distribution using log probability accumulation. A sparse octree structure is used for incremental reads and writes, first updating the occupancy log probability, as shown below: Then, semantic logarithmic confidence accumulation and normalization are performed, as shown below:

[0049] ;

[0050] Where v represents the index of a single voxel in the map. This represents the log-probability of voxel v occupying time t. , Indicates voxel v at time t. The logarithmic odds of occupancy This represents the occupancy observation of voxel v, specifically... Hit status; and Indicates the observational likelihood. This indicates that when a voxel is occupied, an observation is obtained. The conditional probability, This indicates that when a voxel is idle, an observation is obtained. The conditional probability; Represents the prior logarithmic probability; This represents a semantic category index, with categories including pedestrians, vehicles, ground, and obstacles. Voxel representation In category The unnormalized log-log confidence cumulative, This refers to semantic observations, specifically the category confidence scores output by the semantic segmenter. This indicates the conditional probability confidence that the semantic observation comes from the classifier. The normalized class posterior probability is represented by the normalization exponential function. Output, This represents an exponential function with the natural constant as its base. This represents a logarithmic function with the natural constant as its base.

[0051] Example 3, see Figure 1 This embodiment is based on the above embodiment, and the multimodal prediction module specifically includes the following units:

[0052] The multimodal trajectory prediction unit uses a Gaussian mixture model to generate a multimodal future trajectory distribution for each dynamic target, outputting the mode weights, mean trajectory, and covariance, as shown below:

[0053] ;

[0054] in, Indicates the first An index of the tracked target. Representing the future trajectory vector, it represents the positions stacked sequentially in the prediction time domain. This represents the observation context, including the target's historical trajectory, map information, and the observation context vector of its relative position. Indicate target The number of generated modes, Representing modes The weights satisfy and ; Denotes a multivariate Gaussian distribution, indicating the mode. Lower trajectory The conditional distribution with mean is covariance is ; and Representing modes In context The mean locus and covariance matrix under [the given conditions].

[0055] Example 4, see Figure 1 and Figure 3 This embodiment is based on the above embodiment, and the morphology management module specifically includes the following units:

[0056] Establish Markov decision-making processes for high-level task decision-making units. ,by The formal definition of high-level task decisions aims to maximize the cumulative reward of the discount. The action set includes motion objectives and form-switching decisions, represented as follows:

[0057] ;

[0058] in, Represents the strategy, representing the state. Select action The probability distribution; This represents the optimal strategy. Indicating in strategy The mathematical expectation of the trajectory is given below. This indicates the strategy that maximizes the expected discount return. This represents the high-level state at time t. Indicates time The selected action Represents the reward function, Indicates the discount factor;

[0059] Morphological optimization unit, using continuous mixed weights The discrete morphologies are linearly combined into a differentiable hybrid dynamics model. Simultaneously, the morphology optimization objective incorporates the expected task performance, the predicted safety violation probability, and entropy regularization, as shown below:

[0060] ;

[0061] in, A collection representing robot forms, including These represent: wheeled, legged, and crawling; the set size is [size missing]. ; Indicates robot shape index, Representation of form The mixed weights form a vector. , express exist On the probabilistic simplex, satisfying and ; Representation of form The dynamic function is a continuous-time state differential, and the function input is the state. and control vector The output is either the derivative of the current state or the state at the next time step. This represents the hybrid dynamics obtained by linear combination of weights. This represents the control vector, including joint torque and speed commands; Indicated by hybrid power The generated future trajectory samples, This represents the trajectory distribution under a given hybrid dynamics. This represents the trajectory performance function under hybrid dynamics. The expected value is specifically estimated through Monte Carlo sampling. This represents the trajectory performance scoring function. This indicates that there exists a certain moment in the prediction time domain. Make the safety function The probability represents the probability of predicting a safety violation. Describe the control barrier function and define the safety set. , This represents the risk penalty coefficient, used to convert predicted risk into loss. This represents the entropy regularization coefficient, used to encourage sparsity in selection. The negative entropy term indicates a more uniform mixture when it is large, and reducing it can encourage sparsity. This represents the morphological hybrid weight vector obtained through optimization.

[0062] By performing the above operations, this solution addresses the problems of existing multi-modal robot research, which typically treats robot morphology as discrete switching and employs threshold- or rule-based switching strategies. These problems include discontinuous planning and control due to abrupt dynamic changes during switching, high switching jitter and mechanical wear, difficulty in continuously balancing task completion rate, energy consumption, and predicted safety risks, and difficulty in directly integrating dynamic models of different morphologies. This solution linearly combines discrete morphologies into a unified hybrid dynamics using differentiable hybrid weights. In morphology optimization, it introduces safety violation probabilities and entropy regularization from multimodal predictions and learns morphology strategies, making morphology selection a differentiable variable that can be jointly optimized with control. This achieves smooth, low-frequency morphology transitions, reduces switching energy consumption and mechanical shock, and continuously balances task performance and risk constraints, thereby improving task success rate and operational lifespan in multi-modal scenarios. It also facilitates real-time coupling with the lower-level MPC to ensure executability.

[0063] Example 5, see Figure 1 and Figure 4 This embodiment is based on the above embodiment, and the motion control module specifically includes the following units:

[0064] Model predictive control unit, in finite time domain model predictive control Introducing attention weights for multiple targets in the environment The collision probabilities predicted by each target in the multimodal model are then weighted by attention and incorporated into the input. Constraints are imposed to enable real-time motion planning with learnable priorities for high-risk or important targets, as shown below:

[0065] ;

[0066] in, express The number of prediction steps, express Optimized control sequence, Indicates in Predicted system state Indicates the first A reference point for time. , and The cost weights are represented by the state error weight, control cost weight, and terminal state weight, respectively, and are positive semi-definite matrices. The weighting coefficient represents the risk cost, used to adjust the importance of the risk term relative to the trajectory tracking and control costs. Indicates the planned trajectory ,Depend on and Determine, with the target modality trajectory The probability of a collision occurring in the predicted time domain is obtained through the Gaussian overlap approximation, and is a scalar with a value range of [value missing]. ; Indicates attention weights, Indicate the computational objective Score The feature vectors are calculated using relative distance; Indicate target The original scoring function, scalar, The weight vector representing the scoring function, with dimensions and Consistent; This represents the set of safety constraint functions, including minimum distance constraints and velocity boundary constraints. and These represent the upper and lower bounds of the control input components, respectively. Solver usage ;

[0067] The whole-body inverse dynamics unit solves a quadratic programming problem at high frequencies, and... The desired task space acceleration is mapped to joint acceleration or torque, and tracking error and execution cost are minimized under equation dynamics and contact constraints. The output torque command can be directly issued to the driver, as shown below:

[0068] ;

[0069] in, This represents a joint position vector, with the dimension being the number of joints. , Represents the joint velocity vector. Represents the joint acceleration optimization variable. This represents the joint torque vector, which is the final command sent to the drive. The contact force vector is represented by the concatenation of the force vectors at all contact points. This represents the task-space Jacobian matrix, which maps joint accelerations to the task-space acceleration component. Let represent the derivative of the Jacobian matrix with respect to time. express The expected Cartesian acceleration of the output. This represents the task tracking error weight matrix. This represents the torque regularization weight matrix. This represents the historical average torque. Represents the joint space mass matrix. Indicates non-inertial terms, This represents the choice matrix for the controlled degrees of freedom. This indicates contact with Jacobi. Indicates the dimension of the contact constraints. and These represent the upper and lower limits of torque, respectively. Represents the L2 norm, This represents the linearized friction cone constraint.

[0070] By performing the above operations, this solution addresses the problems of existing motion control methods, which typically only perform MPC under a single most probable trajectory or simple obstacle avoidance cost. These problems include the inability to systematically handle multimodal prediction uncertainties in dynamic environments, difficulty in identifying and prioritizing the avoidance of high-risk targets, and the inability to simultaneously satisfy dynamic and contact constraints at the whole-body execution level, leading to conflicts between safety and tracking performance. This solution incorporates the multimodal trajectory probability of each dynamic target into the cost and constraints of MPC after weighting it with feature-based learnable risk attention, and adopts hybrid dynamics to support morphological coupling. Subsequently, high-frequency whole-body inverse dynamics quadratic programming is used to map the desired task acceleration into joint torques that satisfy dynamic and contact constraints, performing safety filtering and online adaptive parameter updates. This achieves prioritized avoidance, safety assurance, and smooth executable control in multi-target, high-uncertainty scenarios.

[0071] Example 6, see Figure 1 and Figure 5 This embodiment is based on the above embodiment, and the control optimization module specifically includes the following units:

[0072] The security filtering unit uses a control barrier function at the front end of the control link to filter the security set. Transformed into a linear constraint on the input. And through real-time secondary planning To obtain the execution input that satisfies the safety constraints by making minimal modifications to the reference control, it is expressed as follows:

[0073] ;

[0074] in, This represents the actual control input executed after being filtered by security, and is a quadratic programming approach. The least squares solution, express The output reference speed, Indicates to of The derivative with respect to natural dynamics, Indicates to of The derivative with respect to control, express Class function, monotonous and , It is a constant;

[0075] Adaptive learning unit, utilizing The optimizer performs small-step online updates to the perception model, dynamic error model, and policy network, as shown below:

[0076] ;

[0077] in, This represents the vector of model parameters being learned. and These represent the model parameter vectors before and after the update, respectively. Indicates the learning rate. This represents the loss function representing the prediction error of the model. This represents the gradient of the loss with respect to the parameters.

[0078] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0079] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0080] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A multi-morphic robot motion control system based on AI perception and decision-making, characterized in that: It includes a perception fusion module, a multimodal prediction module, a morphology management module, a motion control module, and a control optimization module; The perception fusion module integrates data from multiple heterogeneous sensors, outputs an uncertain state estimate with covariance and a dense semantic occupancy map, and simultaneously constructs training data. The multimodal prediction module is used to perform long-term trajectory prediction of dynamic obstacles in the environment and generate multimodal future distributions. The morphology management module performs long-term, multimodal trajectory prediction for dynamic targets in the environment. It uses Gaussian mixture to output modal weights, mean trajectories, and covariance for each target to represent the future uncertainty distribution and to determine the optimal morphology and hybrid dynamic parameters of the robot. The motion control module operates risk attention-driven model predictive control under hybrid dynamics. The attention-weighted multimodal collision probability for each target is incorporated into the cost or constraint. The expected task space acceleration is then mapped to joint torque by high-frequency whole-body inverse dynamics quadratic programming. After satisfying the dynamics and contact constraints, whole-body motion control commands are generated. The control optimization module performs security filtering of underlying instructions and online adaptive updates of system parameters.

2. The multi-morphic robot motion control system based on AI perception and decision-making according to claim 1, characterized in that: The perception fusion module, which integrates multi-source heterogeneous data, outputs a state estimate with covariance and a dense semantic occupancy map, specifically includes the following units: The multimodal data acquisition unit utilizes the data carried on the robot body. Depth camera The robot uses wheel speed sensors to collect raw data in various terrains and records the robot's actual pose and the trajectory of dynamic obstacles in the environment as supervisory labels. The multi-sensor probabilistic fusion unit fuses multi-sensor observations within a Bayesian framework to compute the state posterior; it uses Jacobi linearization for nonlinear observations, outputs the state and covariance, first performs posterior factorization, and then performs extended Kalman filter measurement updates. Occupy grid mapping units, accumulate occupation information through log probability, and maintain voxel semantic distribution by log probability accumulation; use sparse octree structure to perform incremental read and write, first update the occupation log probability; then perform semantic log confidence accumulation and normalization.

3. The multi-morphic robot motion control system based on AI perception and decision-making according to claim 2, characterized in that: The multimodal prediction module specifically includes the following units: The multimodal trajectory prediction unit uses a Gaussian mixture model to generate a multimodal future trajectory distribution for each dynamic target, and outputs the modal weights, mean trajectory, and covariance.

4. The multi-morphic robot motion control system based on AI perception and decision-making according to claim 2, characterized in that: The morphology management module specifically includes the following units: Establish Markov decision-making processes for high-level task decision-making units. ,by The form defines high-level task decisions, with the goal of maximizing the cumulative reward of the discount. The action set includes motion objectives and form switching decisions. Morphological optimization unit, using continuous mixed weights Discrete morphologies are linearly combined into a differentiable hybrid dynamics model, while the morphology optimization objective incorporates the expected performance of the task, the probability of safety violations obtained from the prediction, and entropy regularization.

5. A multi-morphic robot motion control system based on AI perception and decision-making according to claim 1, characterized in that: The motion control module specifically includes the following units: Model predictive control unit, in finite time domain model predictive control The method introduces attention weights for multiple targets in the environment, and incorporates the attention-weighted collision probabilities of each target's multimodal predictions into the algorithm. Constraints enable real-time motion planning with learnable priorities for high-risk or important targets; The whole-body inverse dynamics unit solves a quadratic programming problem at high frequencies, and... The desired task space acceleration is mapped to joint acceleration or torque, and tracking error and execution cost are minimized under equation dynamics and contact constraints. The output torque command can be directly issued to the driver.

6. The multi-morphic robot motion control system based on AI perception and decision-making according to claim 1, characterized in that: The control optimization module specifically includes the following units: The safety filtering unit uses a control barrier function at the front end of the control link to transform the safety set into a linear constraint on the input, and then uses real-time quadratic programming. Make minimal modifications to the reference control to obtain the execution input that satisfies the safety constraints; Adaptive learning unit, utilizing The optimizer performs small-step online updates to the perception model, dynamic error model, and policy network.

Images