Hybrid learning driven trajectory planning method and system for a shop floor robot
By employing a hybrid learning-driven approach, combining the HybridA* algorithm and the conditional diffusion model, the safety, stability, and flexible adaptation issues of robot trajectory planning in smart manufacturing workshops were addressed, achieving smooth, continuous, and high-precision robot trajectories.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- LANZHOU JIAOTONG UNIV
- Filing Date
- 2026-05-21
- Publication Date
- 2026-07-14
AI Technical Summary
In smart manufacturing workshops, existing technologies suffer from poor trajectory smoothness and weak adaptability to dynamic environments due to traditional planning algorithms, insufficient security of deep learning methods, and low inference efficiency of pure diffusion models. These shortcomings make it difficult to meet the requirements of safety, motion stability, and flexible adaptation to different scenarios in robot trajectory planning.
A hybrid learning-driven approach is adopted, combining the HybridA* algorithm to generate global coarse-grained paths, using CNN and MLP to extract environmental features, generating expert trajectories through a conditional diffusion model, and constructing a total energy function for obstacle avoidance, stability, and task constraints to correct the trajectory, thereby achieving safe, stable, and flexible adaptation of the robot trajectory.
It achieves clear global guidance for robot trajectories, smooth and continuous curves, adapts to complex environments, ensures safe obstacle avoidance and precise operation, and improves the docking accuracy of the end-of-line operation and the flexible adaptability to scene changes.
Smart Images
Figure CN122384818A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of mobile robot technology and artificial intelligence, specifically relating to a hybrid learning-driven workshop robot trajectory planning method and system. Background Technology
[0002] As smart manufacturing workshops gradually achieve automation and intelligence, AGVs and mobile operation robots, as core equipment for workshop logistics and operations, need to operate in complex environments with mixed human and machine environments, dense equipment, and frequent dynamic obstacles. This places demands on the safety of trajectory planning, the stability of motion, the accuracy of operation docking, and the flexibility to adapt to different scenarios.
[0003] However, existing technologies have the following shortcomings: First, traditional planning algorithms (such as A*, DWA, and TEB) are highly interpretable and fundamentally safe, but they generate poor trajectory smoothness, making them difficult to adapt to high-dimensional dynamic environments, lacking human-like interactive logic, and having insufficient environmental flexibility. Second, data-driven methods such as deep reinforcement learning and imitation learning rely on data to learn motion strategies, but they have black-box characteristics, making it impossible to guarantee safety in extreme scenarios, and requiring model retraining when workshop constraints change, resulting in poor deployment flexibility. Third, pure diffusion model trajectory generation technology has excellent multimodal fitting capabilities, but low inference efficiency, making it unable to accurately meet the hard constraints of industrial applications such as workshop obstacle avoidance, load stability, and high-precision operations, and thus difficult to directly apply to industrial scenarios.
[0004] Therefore, a new method is urgently needed. Summary of the Invention
[0005] The purpose of this invention is to provide a hybrid learning-driven method and system for workshop robot trajectory planning. This method aims to integrate the global guidance of traditional planning, the multimodal capability of generative models, and the explicit constraint capability of energy functions to solve core problems such as safety assurance, smooth motion, accurate operation, and flexible adaptation of robot trajectory planning in complex workshop environments.
[0006] To achieve the above objectives, this invention provides a hybrid learning-driven workshop robot trajectory planning method and system, comprising the following steps: S1. The robot collects data on the workshop environment, its own dynamics, and the task, generates a local occupancy grid map and a symbolic distance field (SDF), constructs the robot's initial state vector and the target work point state vector, and outputs the above data synchronously to S2, S3, and S5. S2 receives the robot's initial state vector and the target work point state vector output by S1, uses the HybridA* algorithm to generate a global coarse-grained guidance path that satisfies the robot's kinematic constraints, and outputs the path to S3; S3 receives the local occupancy raster map and symbolic distance field SDF output from S1 and the global coarse-grained guidance path output from S2. It extracts environmental image features and path sequence features through CNN and MLP respectively, and obtains conditional features through cross-attention fusion. It uses NMPC to generate expert trajectories, and trains the conditional diffusion model in combination with remote demonstration data. It outputs the conditional features and the trained conditional diffusion model to S4 and S6. S4. Receive the conditional features output by S3 and the robot initial state vector output by S1, sample from the standard Gaussian distribution to generate the inversely denoised initial Gaussian noise trajectory, and output it to S5. S5 receives the initial Gaussian noise trajectory output from S4 and the symbolic distance field SDF, load state, and target work point state vector output from S1, constructs the total energy function including obstacle avoidance, stability, and task constraints, calculates the total trajectory energy, and outputs it to S6. S6: Receive the trained conditional diffusion model output from S3 and the total trajectory energy output from S5, calculate the energy gradient and inject it into the inverse denoising process of the conditional diffusion model to complete the trajectory correction, feed the corrected trajectory back to S5 for iterative iteration, and output the trajectory cluster to S7 after the iteration ends. S7 receives the trajectory cluster output by S6, selects the trajectory with the lowest total energy as the optimal trajectory, and sends it down to the robot's underlying controller.
[0007] Preferably, in S1, the robot's initial state vector is: ; In the formula, This is the robot's initial state vector; This is a Boolean load status identifier, where 0 indicates no load and 1 indicates full load. Let X be the robot's X-axis coordinate in the Cartesian coordinate system of the workshop plane; Let Y be the robot's Y-axis coordinate in the Cartesian coordinate system of the workshop plane; This refers to the robot's heading angle / attitude angle. For the robot's real-time movement speed; The target work point state vector is: ; In the formula, The state vector of the robot's target point; Let X be the X-axis coordinate of the target point in the rectangular coordinate system on the workshop plane. The Y-axis coordinate of the target point in the rectangular coordinate system of the workshop plane; The target point's desired heading angle / attitude angle.
[0008] Preferably, in S3, environmental image features and path sequence features are extracted using CNN and MLP respectively, and conditional features are obtained through cross-attention fusion. The calculation formula for cross-attention fusion is as follows: ; In the formula, This is a cross-attention feature fusion operation; For environmental image features; Features of the path sequence; For multimodal fusion conditional features; The loss function of the conditional diffusion model is expressed as: ; In the formula, Use the loss value to train the model; This represents actual noise. This represents the predicted noise output by the conditional diffusion model. For adding noise to the trajectory; This is for training time steps.
[0009] Preferably, in S4, the initial Gaussian noise trajectory satisfies the following condition: ; In the formula, The initial noise trajectory is t=T; N(0,I) is a standard Gaussian distribution with a mean of 0 and a variance of the identity matrix, and T is the total number of denoising iterations.
[0010] Preferably, in S5, the total energy function is expressed as: ; In the formula, The total energy of the trajectory; For obstacle avoidance energy items; For smooth energy terms; Energy terms constrained for the task; , , These are the weighting coefficients for each energy term.
[0011] Preferably, the obstacle avoidance energy meets the following conditions: When the minimum distance from the trajectory point to the obstacle hour, ;when hour, =0; in, The minimum distance from the trajectory point to the obstacle; This is the safe distance threshold; The obstacle avoidance energy coefficient; To avoid obstacles; Steady energy is represented as: ; In the formula, For stable energy; For continuous trajectory points; For the stable energy coefficient, the value is 1 to 5 under no-load conditions.
[0012] Preferably, in S6, the trajectory correction formula is expressed as: ; In the formula, For the revised Time trajectory; The mean of the trajectory distribution predicted by the diffusion model; This represents the trajectory at the current time step. This is the current iteration step; For conditional features; This represents the variance term predicted by the diffusion model. The energy-guided scale factor; The total energy with respect to the current trajectory The gradient.
[0013] This invention also provides a hybrid learning-driven workshop robot trajectory planning system, comprising: The environmental perception module is used to execute S1, collect workshop environment, self-dynamics and task data, generate local occupancy grid map and symbolic distance field SDF, construct robot initial state vector and target work point state vector, and output the above data synchronously. The global path planning module, connected to the environment perception module, is used to execute S2, receive the robot's initial state vector and the target work point state vector output by the environment perception module, generate a global coarse-grained guidance path that satisfies the robot's kinematic constraints using the HybridA* algorithm, and output the path. The model training module is connected to both the environment perception module and the global path planning module. It is used to execute S3, receive the local occupancy grid map and symbolic distance field SDF output by the environment perception module and the global coarse-grained guiding path output by the global path planning module, extract features and fuse them to obtain conditional features, train the conditional diffusion model, and output the conditional features and the trained conditional diffusion model. The initial noise trajectory sampling module, connected to the model training module, is used to execute S4, receive the conditional features output by the model training module and the robot initial state vector output by the environment perception module, sample and generate the inverse denoising initial Gaussian noise trajectory and output it. The constrained energy calculation module is connected to both the initial noise trajectory sampling module and the environmental perception module. It is used to execute S5, receive the initial Gaussian noise trajectory output by the initial noise trajectory sampling module and the symbolic distance field SDF, load state, and target operation point state vector output by the environmental perception module, construct the total energy function, calculate the total energy of the trajectory, and then output it. The trajectory gradient correction module is connected to both the model training module and the constraint energy calculation module. It is used to execute S6, receive the trained conditional diffusion model output by the model training module and the total trajectory energy output by the constraint energy calculation module, complete the trajectory iterative correction, and output the trajectory cluster. The optimal trajectory output module, connected to the trajectory gradient correction module, is used to execute S7, receive the trajectory clusters output by the trajectory gradient correction module, select the optimal trajectory, and send it to the robot's underlying controller.
[0014] Therefore, the present invention employs the above-mentioned hybrid learning-driven workshop robot trajectory planning method and system, and compared with the prior art, the technical solution of the present invention has the following beneficial effects: (1) The technique of combining HybridA* global coarse-grained path and conditional diffusion model multimodal trajectory generation is adopted to overcome the problems of poor trajectory smoothness and weak dynamic environment adaptability of traditional planning algorithm, as well as the lack of global constraints and easy trapping in local optima of pure diffusion model. It achieves clear global trajectory guidance and smooth and continuous curves, which can be adapted to complex operation scenarios such as mixed human and machine, dense equipment and dynamic obstacles in the workshop.
[0015] (2) Construct an integrated energy function for obstacle avoidance, smoothing, and task constraints. Combine the diffusion model with reverse denoising and energy gradient injection to correct the trajectory. Dynamically adjust the stable energy coefficient through load identification. This overcomes the shortcomings of existing methods that are difficult to balance safe obstacle avoidance, smooth motion, and high operational precision, and cannot adapt to multiple working conditions such as no-load and heavy-load. It can ensure the robot's collision-free safe operation, suppress the shaking of the heavy-load vehicle body, and significantly improve the docking accuracy at the end of the operation.
[0016] (3) The nonlinear model predictive control generates expert trajectories, and the conditional diffusion model is trained by combining manual remote control demonstration data. This overcomes the problems of strong black box nature of deep reinforcement learning and imitation learning, insufficient safety in extreme scenarios, and the need for retraining when the scenario changes. The model has strong generalization ability and does not need to be retrained when the workshop working conditions change. It can adapt to the trajectory planning needs of workshop AGVs and mobile operation robots with high safety and high flexibility.
[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0018] Figure 1A flowchart of an embodiment of the hybrid learning-driven workshop robot trajectory planning method and system of the present invention; Figure 2 This is a schematic diagram of trajectory correction using an energy gradient-guided diffusion model in an embodiment of the hybrid learning-driven workshop robot trajectory planning method and system of the present invention. Figure 3 This is a schematic diagram of the network architecture of the conditional diffusion model (based on one-dimensional time U-Net) of the hybrid learning-driven workshop robot trajectory planning method and system embodiment of the present invention; Figure 4 This is a single-step iterative flowchart of the energy gradient-guided diffusion model trajectory update in an embodiment of the hybrid learning-driven workshop robot trajectory planning method and system of the present invention. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used in the present invention should have the ordinary meaning understood by those skilled in the art.
[0020] Example 1 like Figures 1-4 As shown, this embodiment provides a hybrid learning-driven workshop robot trajectory planning method, applicable to trajectory planning of AGVs (Automated Guided Vehicles) and mobile operation robots in intelligent manufacturing workshops. The parameters, models, and algorithms mentioned in this embodiment are merely examples and do not constitute a limitation of the present invention.
[0021] The hybrid learning-driven workshop robot trajectory planning method and system of the present invention includes the following steps: S1. The robot collects environmental sensor data, its own dynamic state data, and task target data in the workshop to complete environmental characterization and state vector construction.
[0022] The constructed robot state vector, local environment representation data, and task target data are transmitted to steps S2 and S3; at the same time, the collected robot payload state identifier is transmitted to step S5.
[0023] In this step, the robot completes environmental perception and status acquisition. Unlike the traditional method of only collecting spatial information, this step simultaneously incorporates payload status identification, providing a foundation for subsequent stable energy dynamic weighting.
[0024] The robot acquires local environmental observation data of the workshop through LiDAR or depth camera, and generates a local occupancy grid map and SDF (Signed Distance Field) after preprocessing; it collects dynamic information such as robot position, attitude, and velocity, marks load status indicators, and constructs robot state vector; and it parses task instructions to obtain the location of the target work point.
[0025] The robot's state vector is constructed as follows: ; In the formula, This is the robot's initial state vector, representing the robot's current core motion and load states. This is a Boolean load status identifier, where 0 indicates no load and 1 indicates full load. Let X be the robot's X-axis coordinate in the Cartesian coordinate system of the workshop plane; Let Y be the robot's Y-axis coordinate in the Cartesian coordinate system of the workshop plane; This is the robot's heading angle / attitude angle, representing the robot's current orientation; This refers to the robot's real-time movement speed.
[0026] The local environment symbolic distance field (SDF) and the target job point vector are generated and represented as follows: ; In the formula, The state vector of the robot's target point represents the position and posture constraints of the task endpoint; Let X be the X-axis coordinate of the target point in the rectangular coordinate system on the workshop plane. The Y-axis coordinate of the target point in the rectangular coordinate system of the workshop plane; Let be the desired heading angle / attitude angle at the target point, and be the orientation constraint when the robot reaches the destination.
[0027] The robot completes environmental perception and status acquisition at a fixed frequency of 10Hz~50Hz, and simultaneously performs data preprocessing and feature extraction.
[0028] The state data, environmental data, and target data output in this step provide the basic input for subsequent global path planning, condition coding, and energy calculation.
[0029] S2 receives the workshop topology map, robot current status and target work point output by S1, and uses the HybridA* algorithm to generate a global coarse-grained guidance path.
[0030] The generated global coarse-grained guiding path is passed to step S3 as the global conditional input for the conditional diffusion model.
[0031] In this step, the HybridA* algorithm is used to complete the global path search. Unlike traditional path planning, which only generates spatial paths, this step generates a coarse-grained guiding path that adapts to the robot's motion constraints, providing global directional constraints for trajectory generation.
[0032] The robot uses its current position as the starting point of the path and the target work point as the ending point of the path. It loads the workshop topology map, executes the HybridA* path search algorithm, and generates a collision-free discrete waypoint sequence that satisfies the robot's kinematic constraints.
[0033] The global coarse-grained boot path is represented as follows: ; In the formula, Provide a global coarse-grained guidance path for the robot; For the first Waypoint coordinates; This represents the total number of waypoints, ranging from 20 to 100.
[0034] The robot performs path planning at a fixed cycle of 5Hz to 20Hz and updates the global coarse-grained guidance path in real time.
[0035] The global waypoint sequence output in this step provides global path constraint input for subsequent conditional feature encoding; S3 receives the global coarse-grained guidance path output by S2 and the local environmental representation data output by S1, and completes the conditional feature encoding and multimodal feature fusion in the mixed training and inference stages of the conditional diffusion model.
[0036] The fused conditional features are passed to step S4 as constraints for the inverse denoising of the diffusion model.
[0037] In this step, the conditional encoding stage achieves multimodal feature fusion to provide precise constraints for the diffusion model; the model training adopts a hybrid learning strategy and relies on high-precision expert data to complete model convergence; the inference stage achieves trajectory prediction through a one-dimensional temporal convolutional U-Net structure.
[0038] We employ CNN (Convolutional Neural Network) to extract environmental image features from local occupancy raster maps and SDF, and MLP (Multilayer Perceptron) to extract path sequence features from global coarse-grained waypoint sequences. We fuse the two types of features through a cross-attention mechanism and inject the fused features into the one-dimensional temporal convolutional U-Net backbone network.
[0039] The generated fusion condition features are represented as follows: ; In the formula, This is a cross-attention feature fusion operation; For environmental image features; Features of the path sequence; This refers to the conditional features of multimodal fusion.
[0040] The model training employs a hybrid learning strategy, consisting of two steps: Teacher expert data generation and Student diffusion model training. High-precision offline expert trajectory data is generated using the NMPC (Nonlinear Model Predictive Control) optimization algorithm, and a conditional diffusion model is constructed using the Transformer architecture. The parameters of the conditional diffusion model are optimized by training the model with a mixture of expert trajectory data and manual remote control demonstration data, minimizing the mean square error between predicted noise and actual noise.
[0041] The training loss function for constructing the conditional diffusion model is expressed as: ; In the formula, Use the loss value to train the model; This represents actual noise. This represents the predicted noise output by the conditional diffusion model. For adding noise to the trajectory; For training time steps; The fusion conditional features, noise trajectory, and time step encoding are injected into the one-dimensional temporal convolutional U-Net backbone network to output the predicted denoised trajectory, providing input for subsequent energy calculation and trajectory correction.
[0042] This step outputs the pre-trained conditional diffusion model and fusion conditional features, providing the model foundation and constraints for the inverse denoising loop.
[0043] S4 receives the fusion condition features and robot state parameters output by S3, and completes the initial Gaussian noise trajectory sampling for inverse denoising of the diffusion model.
[0044] The initial noise trajectory obtained from the sampling is passed to step S5, and the reverse denoising iteration loop from t=T to t=0 is started.
[0045] In this step, the initial trajectory is sampled from the standard Gaussian distribution. Unlike the traditional trajectory initialization method, this step starts the reverse denoising with conditional features as constraints to ensure the guidance of trajectory generation.
[0046] Based on the preset trajectory length, batch size, and total number of denoising iterations T, initial noisy trajectories are generated by sampling from a standard Gaussian distribution according to the (Batch, Horizon, State_Dim) dimension. Here, Batch is the training / inference batch size, with a value of 8~32; Horizon is the trajectory prediction time domain, with a value of 50~200; and State_Dim is the robot state vector dimension, with a value of 5.
[0047] The initial noise trajectory sampling is represented as follows: ; In the formula, The initial noise trajectory is t=T; N(0,I) is a standard Gaussian distribution with a mean of 0 and a variance of the identity matrix, and T is the total number of denoising iterations.
[0048] Initialize the iteration counter t=T, and start the reverse denoising iteration process at a fixed frequency.
[0049] The initial noise trajectory output in this step provides the initial input for subsequent energy calculations and trajectory correction.
[0050] S5 receives the current iterative trajectory output by S4, the load status identifier, symbolic distance field (SDF), and target work point output by S1, constructs the workshop constraint energy function, and calculates the total trajectory energy.
[0051] The calculated total energy of the trajectory is transferred to step S6 for energy gradient solution.
[0052] In this step, a three-dimensional energy function is constructed that includes obstacle avoidance, stability, and task constraints. Unlike the traditional single energy constraint method, this step introduces a dynamic weighted stable energy term based on load state, which is suitable for heavy-duty handling scenarios in workshops.
[0053] Obstacle avoidance energy is calculated based on the symbolic distance field SDF, stationary energy is calculated based on the second-order difference of the trajectory and the load state, and task energy is calculated based on the error between the trajectory end and the target point. The total trajectory energy is obtained by weighted fusion.
[0054] The total energy function is expressed as: ; In the formula, The total energy of the trajectory; For obstacle avoidance energy items; For smooth energy terms; Energy terms constrained for the task; , , These are the weighting coefficients for each energy term.
[0055] Construct the obstacle avoidance energy function, expressed as: When the minimum distance from the trajectory point to the obstacle hour, ;when hour, =0; in, The minimum distance from the trajectory point to the obstacle; This is the safe distance threshold; The obstacle avoidance energy coefficient; To avoid obstacles.
[0056] Construct a stationary energy function, expressed as: ; In the formula, For stable energy; For continuous trajectory points; For the stable energy coefficient, the value is 1 to 5 under no-load conditions.
[0057] When the load status indicator At that time, the steady-state energy coefficient Increased by 5 times.
[0058] Building Task Energy It is the sum of the Euclidean distance and angle difference between the end of the trajectory and the working machine, used to constrain the alignment accuracy of the end of the trajectory.
[0059] The robot performs calculations of each energy term and merges the total energy at a fixed frequency; it outputs the total energy of the trajectory, providing a constraint basis for subsequent gradient calculation and trajectory correction.
[0060] S6 receives the total trajectory energy and the mean value predicted by the diffusion model from the output of S5, calculates the energy gradient, and completes the inverse trajectory denoising correction.
[0061] Corrected trajectory Feedback is sent to step S5 as input for the next iteration step, until the loop terminates at t=0.
[0062] In this step, the energy gradient is injected into the diffusion model for reverse sampling. Unlike purely data-driven denoising methods, this step forces the trajectory to converge toward a safe and stable region through explicit energy gradient constraints.
[0063] The energy gradient is obtained by differentiating the total energy. Combined with the denoised mean value predicted by the diffusion model, the trajectory is updated and corrected according to the gradient injection formula.
[0064] The trajectory gradient injection correction formula is constructed as follows: ; In the formula, For the revised Time trajectory; The mean of the trajectory distribution predicted by the diffusion model; This represents the trajectory at the current time step. This is the current iteration step; For conditional features; This represents the variance term predicted by the diffusion model. The energy-guided scale factor; The total energy with respect to the current trajectory The gradient.
[0065] The iteration counter is updated round by round (t=t-1), and single-step iteration correction is completed at a fixed frequency.
[0066] This step uses gradient negative feedback correction to ensure that the generated trajectory meets workshop safety and operational constraints.
[0067] S7. Receive the trajectory cluster and corresponding total energy value output after the reverse denoising loop iteration of S4–S6, select the optimal trajectory and send it down to the robot's underlying controller.
[0068] The selected optimal trajectory is directly transmitted to the robot's underlying motion controller, driving the robot to complete the trajectory tracking task.
[0069] In this step, the trajectory is selected based on the total energy value. Unlike the traditional single-index selection method, this step selects the optimal solution by comprehensively considering obstacle avoidance, stability, and task constraints.
[0070] Traverse all feasible trajectories, select the trajectory with the lowest total energy as the optimal trajectory, and parse it into continuous control commands to be sent to the underlying controller.
[0071] The optimal trajectory selection rule is constructed as follows: ; In the formula, The optimal trajectory; The total energy of the trajectory; A mathematical operator for finding the independent variable corresponding to the minimum value of a function; Let X be any candidate trajectory in the trajectory family.
[0072] The trajectory selection, instruction parsing, and issuance operations are completed at a fixed frequency.
[0073] The optimal trajectory output in this step meets all the requirements for safe, stable, and precise operation of the robot in the workshop.
[0074] Therefore, the present invention adopts the above-mentioned hybrid learning-driven workshop robot trajectory planning method and system. This method aims to integrate the global guidance of traditional planning, the multimodal capability of generative models and the explicit constraint capability of energy functions to solve the core problems of robot trajectory planning in complex workshop environments, such as safety assurance, smooth motion, accurate operation and flexible adaptation.
[0075] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A hybrid learning-driven trajectory planning method for workshop robots, characterized in that, Includes the following steps: S1. The robot collects data on the workshop environment, its own dynamics, and the task, generates a local occupancy grid map and a symbolic distance field (SDF), constructs the robot's initial state vector and the target work point state vector, and outputs the above data synchronously to S2, S3, and S5. S2 receives the robot's initial state vector and the target work point state vector output by S1, uses the HybridA* algorithm to generate a global coarse-grained guidance path that satisfies the robot's kinematic constraints, and outputs the path to S3; S3 receives the local occupancy raster map and symbolic distance field SDF output from S1 and the global coarse-grained guidance path output from S2. It extracts environmental image features and path sequence features through CNN and MLP respectively, and obtains conditional features through cross-attention fusion. It uses NMPC to generate expert trajectories, and trains the conditional diffusion model in combination with remote demonstration data. It outputs the conditional features and the trained conditional diffusion model to S4 and S6. S4. Receive the conditional features output by S3 and the robot initial state vector output by S1, sample from the standard Gaussian distribution to generate the inversely denoised initial Gaussian noise trajectory, and output it to S5. S5 receives the initial Gaussian noise trajectory output from S4 and the symbolic distance field SDF, load state, and target work point state vector output from S1, constructs the total energy function including obstacle avoidance, stability, and task constraints, calculates the total trajectory energy, and outputs it to S6. S6: Receive the trained conditional diffusion model output from S3 and the total trajectory energy output from S5, calculate the energy gradient and inject it into the inverse denoising process of the conditional diffusion model to complete the trajectory correction, feed the corrected trajectory back to S5 for iterative iteration, and output the trajectory cluster to S7 after the iteration ends. S7 receives the trajectory cluster output by S6, selects the trajectory with the lowest total energy as the optimal trajectory, and sends it down to the robot's underlying controller.
2. The hybrid learning-driven workshop robot trajectory planning method according to claim 1, characterized in that, In S1, the robot's initial state vector is: ； In the formula, This is the robot's initial state vector; This is a Boolean load status identifier, where 0 indicates no load and 1 indicates full load. Let X be the robot's X-axis coordinate in the Cartesian coordinate system of the workshop plane; Let Y be the robot's Y-axis coordinate in the Cartesian coordinate system of the workshop plane; This refers to the robot's heading angle / attitude angle. For the robot's real-time movement speed; The target work point state vector is: ； In the formula, The state vector of the robot's target point; Let X be the X-axis coordinate of the target point in the rectangular coordinate system on the workshop plane. The Y-axis coordinate of the target point in the rectangular coordinate system of the workshop plane; The target point's expected heading angle / attitude angle.
3. The hybrid learning-driven workshop robot trajectory planning method according to claim 2, characterized in that, In S3, environmental image features and path sequence features are extracted using CNN and MLP respectively, and conditional features are obtained through cross-attention fusion. The calculation formula for cross-attention fusion is as follows: ； In the formula, This is a cross-attention feature fusion operation; For environmental image features; Features of the path sequence; For multimodal fusion conditional features; The loss function of the conditional diffusion model is expressed as: ； In the formula, Use the loss value to train the model; This represents actual noise. This represents the predicted noise output by the conditional diffusion model. For adding noise to the trajectory; This is for training time steps.
4. The hybrid learning-driven workshop robot trajectory planning method according to claim 3, characterized in that, In S4, the initial Gaussian noise trajectory satisfies the following condition: ； In the formula, The initial noise trajectory is t=T; N(0,I) is a standard Gaussian distribution with a mean of 0 and a variance of the identity matrix, and T is the total number of denoising iterations.
5. The hybrid learning-driven workshop robot trajectory planning method according to claim 4, characterized in that, In S5, the total energy function is expressed as: ； In the formula, The total energy of the trajectory; For obstacle avoidance energy items; For smooth energy terms; Energy terms constrained for the task; , , These are the weighting coefficients for each energy term.
6. The hybrid learning-driven workshop robot trajectory planning method according to claim 5, characterized in that, Obstacle avoidance energy must meet the following conditions: When the minimum distance from the trajectory point to the obstacle hour, ;when hour, =0; in, The minimum distance from the trajectory point to the obstacle; This is the safe distance threshold; The obstacle avoidance energy coefficient; To avoid obstacles; Steady energy is represented as: ； In the formula, For stable energy; For continuous trajectory points; For a stable energy coefficient, the value is 1 to 5 under no-load conditions.
7. The hybrid learning-driven workshop robot trajectory planning method according to claim 6, characterized in that, In S6, the trajectory correction formula is expressed as: ； In the formula, For the revised Time trajectory; The mean of the trajectory distribution predicted by the diffusion model; This represents the trajectory at the current time step. This is the current iteration step; For conditional features; This represents the variance term predicted by the diffusion model. The energy-guided scale factor; The total energy with respect to the current trajectory The gradient.
8. A hybrid learning-driven workshop robot trajectory planning system, applied to the hybrid learning-driven workshop robot trajectory planning method according to any one of claims 1-7, characterized in that, include: The environmental perception module is used to execute S1, collect workshop environment, self-dynamics and task data, generate local occupancy grid map and symbolic distance field SDF, construct robot initial state vector and target work point state vector, and output the above data synchronously. The global path planning module, connected to the environment perception module, is used to execute S2, receive the robot's initial state vector and the target work point state vector output by the environment perception module, generate a global coarse-grained guidance path that satisfies the robot's kinematic constraints using the HybridA* algorithm, and output the path. The model training module is connected to both the environment perception module and the global path planning module. It is used to execute S3, receive the local occupancy grid map and symbolic distance field SDF output by the environment perception module and the global coarse-grained guiding path output by the global path planning module, extract features and fuse them to obtain conditional features, train the conditional diffusion model, and output the conditional features and the trained conditional diffusion model. The initial noise trajectory sampling module, connected to the model training module, is used to execute S4, receive the conditional features output by the model training module and the robot initial state vector output by the environment perception module, sample and generate the inverse denoising initial Gaussian noise trajectory and output it. The constrained energy calculation module is connected to both the initial noise trajectory sampling module and the environmental perception module. It is used to execute S5, receive the initial Gaussian noise trajectory output by the initial noise trajectory sampling module and the symbolic distance field SDF, load state, and target operation point state vector output by the environmental perception module, construct the total energy function, calculate the total energy of the trajectory, and then output it. The trajectory gradient correction module is connected to both the model training module and the constraint energy calculation module. It is used to execute S6, receive the trained conditional diffusion model output by the model training module and the total trajectory energy output by the constraint energy calculation module, complete the trajectory iterative correction, and output the trajectory cluster. The optimal trajectory output module, connected to the trajectory gradient correction module, is used to execute S7, receive the trajectory clusters output by the trajectory gradient correction module, select the optimal trajectory, and send it to the robot's underlying controller.