Robotic motion planning method, system, medium, and robot

Abstract

The application provides a robot motion planning method, system, medium and robot, receives a planning request code from an upper application to form a planning condition c, samples K latent variables z from a preset probability distribution, outputs K candidate trajectories through a generative planning network, and performs constraint verification; the candidate trajectories passing the constraint verification are scored and sorted, one or more candidate trajectories with the highest score are selected as local optimization initial values, a limited number of optimization, time parameterization or collision margin correction is performed, if the number of candidate trajectories passing the constraint verification is insufficient or local optimization fails, a planner rollback is triggered to write an active oversampling pool, which is used for incremental training of the generative planning network. Through the construction of a closed-loop system of "generation-verification-backup-learning", the application improves the online planning speed and success rate under the premise of ensuring the planning safety and robustness, and enables the system to have the ability of continuous self-optimization.

Description

Technical Field

[0001] This invention relates to the field of robotics, and more specifically, to a robot motion planning method, system, medium, and robot. Background Technology

[0002] Robot motion planning aims to generate a collision-free, executable trajectory for a robot from an initial state to a target state, satisfying various constraints. Traditional motion planning methods, such as sampling methods based on fast expanding random trees or trajectory optimization methods, while theoretically highly reliable, often suffer from problems such as excessively long planning times, sensitivity to initial values, and susceptibility to local optima when dealing with robotic arms with redundant degrees of freedom, narrow spaces with dense obstacles, or complex scenarios requiring selection from multiple feasible solutions. These issues make it difficult to meet the high efficiency and real-time requirements of industrial scenarios.

[0003] To improve planning speed, several methods employing machine learning models have emerged in recent years to accelerate motion planning. These methods typically learn from a large number of planning cases to quickly generate one or more candidate trajectories or joint configurations, which are then validated and selected. However, these methods still have significant drawbacks: First, their planning process is inherently open-loop. When all candidate trajectories generated by the model are rejected due to failure to meet hard engineering constraints such as collisions and joint limitations, the system will fail to plan, lacking necessary safety guarantees and robustness. Second, these methods typically lack the ability to learn online and improve themselves. For difficult scenarios not fully covered in the training data, the low planning success rate persists, and performance cannot be adaptively improved through interaction with the environment in actual deployment. Therefore, current technologies have not yet provided a comprehensive solution that balances planning speed, safety, robustness, and continuous evolution capabilities. Summary of the Invention

[0004] In view of the deficiencies in the prior art, the purpose of this invention is to provide a robot motion planning method, system, medium, and robot.

[0005] A robot motion planning method provided by the present invention includes: Step S10: The system receives a planning request from the upper-layer application. The planning request is a fetching task and includes corresponding planning information. Step S20: The received planning request is encoded into planning condition c by the condition encoding module, and K latent variables z are sampled from a preset probability distribution; Step S30: Using a generative programming network, output K sets of candidate trajectories in parallel based on the planning condition c and K latent variables z; Step S40: Perform constraint verification on the K groups of candidate trajectories and filter out candidate trajectories that fail the constraint verification; Step S50: Score and rank the candidate trajectories that have passed the constraint verification; Step S60: Select one or more candidate trajectories with the highest scores as the initial values ​​for local optimization, and perform a finite number of optimizations, time parameterization, or collision margin correction; Step S70: If the number of candidate trajectories that pass the constraint verification is insufficient or local optimization fails, the planner is triggered to back off, and the failure conditions, back-off results and final execution results are written into the active sampling pool. Step S80: Output the final executable trajectory and record the planning time, pass rate, failure type, optimization residual and safety verification results for incremental training of the generative planning network.

[0006] Furthermore, the planning information includes: current joint state, target pose / path, obstacles, constraint parameters, and task preferences.

[0007] Furthermore, the constraint verification includes filtering candidate trajectories that violate end-effector accuracy, joint limit, collision, self-collision, velocity / acceleration, and smoothness constraints.

[0008] Furthermore, the scoring ranking includes: ranking multi-objective scores according to error, collision margin, trajectory length, energy consumption, distance from the current attitude, and task preference.

[0009] A robot motion planning system provided by the present invention includes: Module M10: The system receives a planning request from an upper-layer application. The planning request is a fetching task and contains corresponding planning information. Module M20: Encodes the received planning request into planning condition c through the condition coding module, and samples K latent variables z from a preset probability distribution; Module M30: Outputs K sets of candidate trajectories in parallel based on the planning condition c and K latent variables z through a generative programming network; Module M40: Performs constraint verification on K groups of candidate trajectories and filters out candidate trajectories that fail the constraint verification; Module M50: Scores and sorts the candidate trajectories that have passed the constraint verification; Module M60: Select one or more candidate trajectories with the highest scores as initial values ​​for local optimization, and perform a finite number of optimizations, time parameterization, or collision margin correction; Module M70: If the number of candidate trajectories that pass the constraint verification is insufficient or local optimization fails, the planner is triggered to back off, and the failure conditions, back-off results and final execution results are written into the active sampling pool. Module M80: Outputs the final executable trajectory and records the planning time, pass rate, failure type, optimization residual, and safety verification results for incremental training of the generative planning network.

[0010] Furthermore, the planning information includes: current joint state, target pose / path, obstacles, constraint parameters, and task preferences.

[0011] Furthermore, the constraint verification includes filtering candidate trajectories that violate end-effector accuracy, joint limit, collision, self-collision, velocity / acceleration, and smoothness constraints.

[0012] Furthermore, the scoring ranking includes: ranking multi-objective scores according to error, collision margin, trajectory length, energy consumption, distance from the current attitude, and task preference.

[0013] According to the present invention, a computer-readable storage medium storing a computer program is provided, wherein when the computer program is executed by a processor, the steps of the robot motion planning method are implemented.

[0014] A robot according to the present invention includes the aforementioned robot motion planning system.

[0015] Compared with existing technologies, this invention integrates generative multi-candidate models, trajectory-level constraints, coverage feedback, local optimization, and traditional planner backoff into a continuously improving planning acceleration closed loop, achieving the following beneficial effects: (1) Improved online planning speed: By outputting K sets of candidate solutions or candidate trajectories in parallel through the conditional generative planning network, high-quality initial values ​​can be provided before the traditional sampling / optimization planner starts a large-scale search, reducing search time and the number of local optimization iterations.

[0016] (2) Improved candidate diversity and coverage: The latent variable z drives the generation of multi-branch candidates under the same planning conditions, enabling the system to cover multiple feasible solution branches of redundant robotic arms and reduce the probability of failure in difficult scenarios caused by mode collapse.

[0017] (3) Improved executability at the trajectory level: This scheme incorporates the current state, target path, environmental constraints, velocity / acceleration, continuity and smoothness into candidate generation, training loss and online verification, so that the network output is not limited to single-point IK solution, but is closer to the executable motion planning result.

[0018] (4) Improved safety and engineering stability: Neural network candidates must be verified, sorted and optimized locally before they can be output; when there are insufficient candidates or optimization fails, the traditional planner is triggered to back off, so the existing robot safety planning link can be preserved while accelerating.

[0019] (5) Improved adaptability to difficult areas: The coverage feedback active sampling mechanism identifies low coverage areas based on candidate pass rate, failure type, backoff frequency and optimized residuals, and generates high-value samples in a targeted manner, so that the model can continuously improve near obstacles, near joint limits, singular areas and historical failure scenarios.

[0020] (6) Improved local optimization success rate: Using the candidates output by the network as seeds for the local optimizer can reduce the sensitivity of the traditional optimizer to the initial value, and select the initial value with greater collision margin, shorter trajectory or smoother path by sorting multiple candidates.

[0021] (7) Improved training resource utilization: Compared with random expansion of the training set, this scheme prioritizes supplementing samples in regions with low coverage, high failure rate, and high engineering value, which can reduce the collection of invalid samples and training time, and enable the model to reach the target coverage rate more quickly.

[0022] (8) Improved system pluggability: Generative planning network can be connected as a front-end accelerator of existing planning stack without replacing collision detection, trajectory optimization, time parameterization and safety monitoring modules; therefore, it is easy to deploy in existing robot control systems and can switch models or backoff strategies according to tasks.

[0023] (9) Quantifiable verification indicators include: average planning time and 95th percentile time, candidate coverage, end position / attitude error, trajectory smoothness, local optimization residual, backoff trigger rate, backoff successful sample reuse rate, and improvement in the pass rate of difficult areas. Attached Figure Description

[0024] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the overall process flow of the method provided in the embodiments of this application; Figure 2 A schematic diagram of the coverage feedback active resampling mechanism provided in the embodiments of this application; Figure 3 A flowchart of a robot motion planning method provided in an embodiment of this application; Figure 4 A schematic diagram of the hardware architecture of a robot motion planning system provided in this application embodiment; Figure 5 This is a timing diagram of the signaling interaction between various modules in a robot motion planning method provided in an embodiment of this application. Detailed Implementation

[0025] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0026] Example 1 This embodiment provides a specific implementation of a robot motion planning method and system based on generative models and feedback optimization. In this embodiment, a typical collaborative robot obstacle avoidance and grasping task is used as an example to illustrate the complete process of the technical solution of this application.

[0027] Please see Figure 4 This figure illustrates a hardware architecture diagram of a robot motion planning system provided in an embodiment of this application. The system can be deployed in a computing unit 300, which can be a high-performance industrial control computer or an embedded motherboard with a graphics processing unit. The computing unit 300 internally includes at least one processor 310 and a memory 320. The processor 310 is used to execute instructions and programs stored in the memory 320. The memory 320 stores software modules for implementing the method of this application, specifically including: a first planning module 321, a verification and evaluation module 322, a second planning module 323, and a model update module 324. The computing unit 300 communicates with a robot controller 400 via a high-speed industrial bus such as EtherCAT or Profinet to send the calculated executable trajectory. Simultaneously, the computing unit 300 is also connected to one or more sensors 500, such as one or more depth cameras, LiDAR, or torque sensors, to obtain real-time information about the robot's environment. In this embodiment, the robot is a redundant collaborative robotic arm with 7 degrees of freedom, and the sensor 500 is a depth camera fixed above the work unit, which is used to capture point cloud data of the work surface and the surrounding environment.

[0028] Please refer to the following: Figure 1 and Figure 3 , Figure 1 This is a schematic diagram of the overall process of a robot motion planning method and system provided in an embodiment of this application. Figure 3 A flowchart illustrating a robot motion planning method provided in this embodiment. The complete working process of this embodiment is as follows: First, in step S10 (reading the planning request), the system receives a planning request from an upper-layer application (e.g., a warehouse management system or a production line control system). In this embodiment, the planning request is specifically a grasping task, which contains specific and rich planning condition information. Specifically, the planning conditions include at least: the robot's current state, i.e., the current angle values ​​of the seven joints; the target state of the task, i.e., the target grasping pose that the robot's end effector needs to reach, which is usually represented by a homogeneous transformation matrix containing position (x, y, z) and orientation (quaternions or Euler angles); and environmental information of the robot's environment. This environmental information consists of real-time 3D point cloud data collected by sensors (depth cameras). It should be noted that the point cloud data is usually preprocessed before being sent to the planning system, such as denoising, downsampling, and conversion into a representation more suitable for collision detection, such as a voxel grid or a symbolic distance field.

[0029] Subsequently, in step S20 (conditional encoding and sampling), the conditional encoding module located in the computing unit 300 processes the received planning request. The conditional encoding module encodes input information of different forms and dimensions into a single planning condition c with a fixed dimension. Specifically, it concatenates or fuses the robot's initial joint angle, target end-effector pose, and environmental point cloud information after feature extraction (e.g., global environmental features extracted through a small convolutional neural network) to form a planning condition c that comprehensively describes the current planning problem. This encoding process ensures that the subsequent planning model can handle various planning tasks. Simultaneously, the system randomly samples K latent variables z from a preset probability distribution (e.g., a standard multidimensional normal distribution). In this embodiment, the value of K can be an integer greater than 1, such as K=50. It can be understood that each latent variable z can be considered a "seed of creativity" to guide the generative model to generate different candidate solutions.

[0030] Subsequently, in step S30 (generating candidate trajectories), the first planning module 321, specifically in this embodiment, is a pre-trained conditional variational autoencoder (CDE) serving as a generative planning network, receiving the planning condition c and K latent variables z output by the conditional encoding module. This CDE is trained to learn the mapping relationship from the planning conditions and latent variables to feasible trajectories. Using its decoder, it decodes each pair (c, z_i) (where i ranges from 1 to K) in parallel and non-iteratively into a candidate joint space trajectory. Therefore, in a very short time (typically on the order of tens of milliseconds), this module can generate K=50 sets of candidate trajectories at once. Each candidate trajectory is a time series defining a series of intermediate joint angles from the initial joint angle to near the target joint angle.

[0031] Subsequently, in step S40 (constraint verification and filtering), the verification and evaluation module 322 performs rigorous and deterministic constraint verification on the K=50 candidate trajectories generated by the first planning module 321. This is a crucial safety step to ensure that only physically and engineeringally feasible trajectories can proceed to the next stage. In this embodiment, the constraint verification includes, but is not limited to, the following checks: 1. Collision detection: checking whether the robot's link model collides with the environment model constructed from point cloud data at each time point of the trajectory execution; 2. Self-collision detection: checking whether collisions occur between different links of the robot; 3. Joint limit check: checking whether all joint angles in the trajectory are within the physical range of motion of each joint of the robot; 4. End-effector pose error check: checking whether the error between the end-effector pose corresponding to the endpoint of each candidate trajectory and the target pose given in the planning request is within a preset tolerance range (e.g., position error less than 1 mm, attitude error less than 0.5 degrees). Any candidate trajectory that fails any of the above checks will be filtered by the verification and evaluation module 322.

[0032] Subsequently, the system will proceed with different processing steps depending on whether there are qualified candidate trajectories that pass all constraint checks. For details, please refer to [link to relevant documentation]. Figure 5 The signaling interaction timing diagram is shown.

[0033] The first scenario is successful verification. Assume that in this planning process, 10 out of K=50 candidate trajectories pass all constraint verifications. In this case, the process proceeds to step S50 (candidate sorting). The verification and evaluation module 322 sorts these 10 qualified candidate trajectories according to one or more preset evaluation indicators. These indicators may include: trajectory length (preferring shorter paths), trajectory smoothness (preferring trajectories with smaller changes in joint velocity and acceleration), energy consumption (preferring trajectories with lower estimated energy consumption), or collision margin (preferring trajectories farthest from obstacles). In this embodiment, the system uses a weighted sum comprehensive scoring function for sorting.

[0034] As a preferred implementation, in step S60 (local optimization), the system can select one or more candidate trajectories with the highest scores and use them as initial values ​​(or "seeds") for the local optimization module. The local optimization module can be a trajectory optimization-based planner, such as the TrajOpt or CHOMP algorithm. Given that the initial trajectory provided by the generative model is already very close to a feasible solution, the local optimizer typically only needs a few iterations to further refine the trajectory, making it smoother, more optimal, and strictly satisfying all constraints. The optimized trajectory will be sent to the robot for execution via the computing unit 300 and the robot controller 400 as the final executable trajectory. This process corresponds to... Figure 5The "Validation Successful" branch shown in the diagram involves the application layer initiating a planning request (planRequest()), which in turn calls the candidate generation function (generateCandidates()) of the first planning module. After successful internal validation (validate()), the system directly returns the trajectory (returnTrajectory()) to the application layer.

[0035] The second scenario is a failure to pass the constraint check. Suppose that because the target object is located in a confined space with a complex structure, the K=50 candidate trajectories generated by the first planning module 321, although diverse, all fail the constraint check due to collisions with the environment. In this case, the validation and evaluation module 322 determines that the number of qualified candidate trajectories is zero. This indicates that the first planning model has failed in the current specific scenario.

[0036] Accordingly, the system will trigger a safety rollback mechanism to execute step S70 (rollback and resampling). For example... Figure 1 and Figure 5 As shown, the control flow will then shift to the second planning module 323, which is the traditional planner fallback module. In this embodiment, this module is implemented as a classic planner based on the Rapidly Expanding Random Tree (RRTConnect) bidirectional connection algorithm. This planner is known for its high reliability and probabilistic completeness. Although its planning speed is usually slower than generative models, it has a high probability of finding a solution within a sufficient time, provided a solution exists. The second planning module 323 receives the exact same planning request as before and begins searching in the state space. Assuming that it successfully finds an executable obstacle avoidance trajectory that satisfies all constraints after approximately 500 milliseconds, this trajectory will be the final result of this planning task and will be immediately output as an executable trajectory, thus ensuring the successful completion of the task and the overall stability of the system. This process corresponds to... Figure 5 The "Validation Failure" branch shown in the diagram involves the planning system calling the fallback planning function (fallbackPlan()) of the second planning module after the internal validation (validate()) fails. After obtaining the fallback trajectory, it returns to the application layer to ensure the availability of the service.

[0037] Meanwhile, the failure event itself constitutes a valuable learning signal. The model update module 324 records the planning condition c (including initial / target state and environmental information) that caused the first planning model to fail, as well as the success trajectory generated by the second planning module 323. This (failure condition, success trajectory) data pair is considered a "high-value sample" and is stored in a database called the "failure sample pool".

[0038] Finally, in step S80 (output and recording), the trajectory generated through either the first or second scenario will be output and executed. Simultaneously, the system will record a detailed log of this planning process, including whether the first or second planning model was used, the total time consumed, and the pass rate of candidate trajectories. These statistics will be used to evaluate the online planning performance of the first planning model.

[0039] When the system is idle, or according to a preset cycle (e.g., every night), the model update module 324 will initiate the offline incremental training process, which is the function of the active imputation and incremental training module. This module will preferentially extract these recorded high-value difficult samples from the failure sample pool and use them to fine-tune the conditional variational autoencoder model in the first planning module 321. In this way, the model is guided to learn how to generate the correct trajectory in previously failed difficult scenarios.

[0040] Through this embodiment, the technical solution of this application constructs a complete closed loop of "generation-verification-rollback-learning". In most conventional scenarios, the system utilizes the first planning model to achieve rapid planning, significantly improving efficiency. In occasional difficult scenarios, the rollback mechanism of the second planning model ensures the success rate and safety of the task. More importantly, the system can learn from failures, and through active sampling and incremental training, the planning success rate of the first planning model in similar difficult scenarios continuously improves, thereby achieving continuous self-optimization and evolution.

[0041] Example 2 This embodiment is a variant of Embodiment 1, intended to illustrate the universality of the technical solution of this application. Its core idea does not depend on any specific generative model. The overall system architecture and hardware configuration of this embodiment (e.g.) Figure 4 As shown, the functions of all modules except the first planning module 321 are consistent with those in Embodiment 1. The core difference in this embodiment is that the generative planning network in the first planning module 321 is replaced with a more advanced conditional diffusion model.

[0042] Conditional diffusion models are powerful generative models that generate data through a simulated "backward diffusion" process. Specifically, their operation can be understood in two stages: Forward pass (training phase): During the training phase, the model learns how to progressively and iteratively add Gaussian noise to clean, realistic trajectory data until the trajectory data becomes purely random noise. The number of steps in this process is fixed, for example, N=1000 steps.

[0043] Reverse Process (Generation Phase): In step S30 (generating candidate trajectories) of online planning, the model performs a reverse process when trajectories need to be generated. Specifically, the first planning module 321 first generates K (e.g., K=50) initial trajectory tensors composed entirely of random Gaussian noise. Then, in an iterative denoising loop, the conditional diffusion model, guided by the planning condition c, predicts and subtracts the noise that should have been added in the forward process at each step. This denoising process is performed in parallel on the K noise tensors for N steps (to accelerate this, the number of steps during inference can be much smaller than the 1000 steps during training, e.g., N=20 steps). After N iterations, the K initial random noise tensors are gradually "sculpted" into K candidate joint trajectories with clear structures and highly correlated with the planning condition c.

[0044] Compared to the conditional variational autoencoder used in Example 1, using the conditional diffusion model as the first programming model may offer the following potential advantages: 1. Higher quality generation: Diffusion models have been proven in fields such as image generation to generate extremely detailed and high-quality samples. In the field of trajectory planning, this means that the initial candidate trajectories they generate typically have better intrinsic smoothness and temporal continuity, more closely resembling a physically natural motion.

[0045] 2. Higher pass rate: Due to the higher quality of the initial trajectories, they are likely to pass a higher percentage of the checks in step S40 (constraint verification filtering), especially the velocity / acceleration limit and trajectory continuity check.

[0046] 3. Better results as initial value for optimization: A smoother and more natural initial trajectory, when used as the initial value for the local optimization module in step S60, can often enable the optimization process to converge to a high-quality solution more quickly, and may even reduce the need to trigger local optimization.

[0047] The subsequent steps in this embodiment include constraint verification, evaluation and ranking (S40, S50), optional local optimization (S60), a fallback mechanism in case of failure (S70), and closed-loop learning based on failed samples. The logic and process are consistent with those in Embodiment 1. If all K=50 candidate trajectories generated by the diffusion model fail the verification, the system will also activate the second planning module 323 (e.g., RRTConnect) for fallback planning, and store the failed planning conditions and the successfully fallback trajectories in the failure sample pool for subsequent incremental training of the conditional diffusion model.

[0048] This embodiment fully demonstrates the flexibility and scalability of the dual-model planning and closed-loop learning framework proposed in this application. The first planning module 321 can be regarded as a "pluggable" component. Regardless of whether a conditional variational autoencoder, a conditional diffusion model, or other types of generative models are used, such as a conditional normalized flow model or a Transformer-based sequence model as described in Embodiment 4, the core advantages of the entire system—namely, balancing speed, security, and continuous learning capability—are maintained. This provides the freedom to select the most suitable generative model based on specific application scenarios and available computing resources, and supports the broader scope of protection of this application.

[0049] Example 3 This embodiment aims to provide a more detailed explanation of the "coverage feedback and active supplementation" mechanism of this application. This mechanism is key to achieving continuous self-evolution of the model, corresponding to... Figure 1 The active imputation and incremental training modules in the middle and Figure 2 The detailed process is shown below. This embodiment enhances and refines the functionality of the model update module 324 based on the system of Embodiment 1.

[0050] Please see Figure 2 The figure illustrates in detail the working principle of the coverage feedback active sampling mechanism. The model update module 324 contains a crucial "coverage evaluation unit." The core task of this unit is to continuously monitor and analyze the online planning performance of the first planning module 321, thereby accurately locating its "knowledge blind spots" or "capability gaps," i.e., low-coverage areas.

[0051] The working process of this embodiment can be described as follows: Assume that the planning system of this application has been deployed and running online for one week on an industrial assembly line. During this period, the system processed a large number of planning requests, for example, 100,000. For each planning request, the system records detailed log information in step S80 (output and recording). This information includes at least: the input planning condition c (or its key features), the verification results of the K candidate trajectories generated by the first planning model (i.e., how many passed and for what reasons they failed), whether the second planning model was ultimately triggered to backtrack, and the total planning time, etc. This massive amount of online data forms the basis for coverage evaluation.

[0052] During the offline analysis phase (e.g., on a server or when the system is idle), the coverage evaluation unit in the model update module 324 begins its work. It first needs to partition the robot's state space or workspace in some form to define different "regions." In this embodiment, a specific and effective method is to divide the workspace accessible to the robot's end effector into a three-dimensional grid map, for example, each grid cell being 10 cm x 10 cm x 10 cm in size.

[0053] Then, the coverage evaluation unit iterates through the 100,000 planning logs recorded that week, classifying each planning task into a corresponding grid based on the main region where its target pose or trajectory is located. Next, it calculates the key performance indicators for each grid, the most crucial of which is the "planning success rate" (or "pass rate") of the first planning model. The success rate can be defined as the proportion of all planning tasks falling into that grid where the first planning model can generate at least one candidate trajectory that passes constraint validation.

[0054] Through such statistical analysis, the system may discover that although the average success rate of the first planning model is over 95% in most areas, in a specific area behind the robot base, near a large device (e.g., composed of several adjacent grids), the success rate drops sharply to only 15%. Furthermore, by analyzing the failure logs, the system can also identify that the main reasons for failure in this area are self-collision and the sixth joint (J6) reaching its joint limit.

[0055] At this point, areas with success rates significantly below the average are automatically identified by the system as "low-coverage areas" and placed in a high-priority low-coverage area queue. The criteria for identifying low-coverage areas can be that their success rate is below a preset threshold (e.g., below 50% or below half the average success rate).

[0056] After identifying the weak area, the system will supplement the samples accordingly. The model update module 324 will initiate a process to generate high-value samples. This process will proactively and programmatically generate hundreds of new, challenging planning tasks in and around the newly identified low-coverage area. For example, it will randomly sample hundreds of new target poses within that 15% success rate area.

[0057] For these newly generated "targeted" tasks, the system invokes a reliable second planning module 323 (e.g., an optimization-based planner TrajOpt, which can find high-quality solutions given sufficient time) to generate high-quality, collision-free, self-collision-free success trajectory samples that satisfy all constraints. These newly generated (planning conditions, success trajectories) data pairs are specifically created to compensate for the shortcomings of the first planning model and are therefore referred to as "high-value samples".

[0058] Finally, these high-value samples will be added to the training dataset. In the next round of model training or updates, the system will employ a priority training strategy. For example, through a "priority experience replay" mechanism, the probability of selecting these high-value samples from low-coverage areas is significantly increased when batches of data are drawn from the training dataset for training. This ensures that the model can focus on and learn to solve these previously difficult problems when updating parameters. After such targeted reinforcement training, the parameters of the first planning model are updated (network update), and its shortcomings are effectively compensated for.

[0059] It is foreseeable that, in this embodiment, after such a round of active sampling and incremental training, when the system is redeployed, the new model's planning success rate in the previously identified difficult area is expected to significantly increase from 15% to over 80%. This will directly lead to a substantial reduction in the trigger frequency of the second planning model's rollback, thereby improving the average planning efficiency and operational stability of the entire system, truly realizing the system's continuous learning and adaptive evolution.

[0060] Example 4 This embodiment aims to demonstrate that the technical solution of this application is not only applicable to point-to-point planning tasks (such as the grasping task described in Embodiment 1), but can also be efficiently applied to more complex trajectory-level planning tasks, such as path following tasks such as welding, grinding, and gluing commonly used in industrial applications.

[0061] In this type of task, the content and form of the planning request differ from those in the aforementioned embodiments. Here, the planning request is no longer a single target pose, but rather a desired path defined in a Cartesian workspace. This path is typically described by a sequence of N pose points (e.g., N=100), defining the trajectory and posture that the robot's end effector (such as a welding torch or grinding head) needs to precisely follow.

[0062] To adapt to this sequence-to-sequence planning task, this embodiment adopts a model architecture more suitable for processing sequence data for the generative planning network in the first planning module 321. Specifically, a Transformer-based sequence-to-sequence model can be used. The encoder part of this model is responsible for receiving and understanding the entire sequence of Cartesian path points as input, encoding it into a context vector rich in path geometry and temporal information. The decoder part then uses this context vector and the starting joint angle as conditions to autoregressively or in parallel generate the corresponding sequence of joint space trajectory points, which also contains N time points. By introducing a latent variable z into the input, or by using different sampling strategies during the decoding process, this Transformer model can also generate K groups (e.g., K=10) of candidate complete joint space trajectories at once.

[0063] The workflow of this embodiment also follows... Figure 3 The steps are shown, but some key steps are emphasized: 1. In steps S20 (conditional encoding and sampling) and S30 (generating candidate trajectories), as described above, the system processes sequence data and uses the Transformer model to generate K sets of complete joint space trajectory sequences.

[0064] 2. In step S40 (constraint verification filtering), the checks performed by the verification evaluation module 322 become more complex and stringent. In addition to routine collision detection, self-collision detection, and joint limit checks, the following constraints related to the path-following task must be specifically checked: – Path tracking error: The generated candidate joint trajectory is transformed back to Cartesian space through positive kinematics, and its position and orientation deviation from each corresponding point on the desired welding path is calculated. The candidate trajectory is considered qualified only if the maximum or average deviation of the entire path is within the preset process tolerance (e.g., position deviation less than 0.5 mm).

[0065] –Trajectory Smoothness and Continuity: This checks whether the velocity, acceleration, and even jerk of the generated joint trajectory are continuous and within the limits allowed by robot dynamics. In applications such as welding or gluing, smooth motion is crucial for ensuring process quality. Any trajectory with sudden velocity changes or excessive acceleration will be filtered out.

[0066] 3. In step S50 (candidate sorting), in addition to collision margin, the sorting criteria are more importantly the accuracy and smoothness of path tracking. The system will prioritize the candidate trajectory with the smallest tracking error and the most stable joint velocity changes.

[0067] 4. In step S60 (local optimization), the local optimization module plays a more crucial role. Even if the candidate trajectories generated by the first planning model largely meet the requirements, fine optimization is usually necessary to achieve extremely high process accuracy. An efficient approach is as follows: if the tracking errors of all K candidate trajectories slightly exceed the tolerance, the system can select the one with the smallest error and use it as the initial value for an optimization-based planner (e.g., TrajOpt). Then, the optimizer optimizes the trajectory with path tracking accuracy as a hard constraint and minimizing joint motion or time as the objective. Because of a very good initial solution, this optimization process is usually very fast.

[0068] 5. In step S70 (backoff and supplementation), the triggering condition for the backoff mechanism can be: the tracking error of all K candidate trajectories is too large, or the local optimization process fails to converge. At this time, the system can call a more time-consuming but more powerful second planning module 323, such as an optimization algorithm specifically designed for path planning, to plan from scratch or with a coarse interpolated trajectory as the initial value.

[0069] The learning and updating mechanism (model update module 324) described in this embodiment also applies. The system will specifically record path features that cause large tracking errors, such as corners with large curvature changes and road sections with drastic attitude changes. These scenes will be identified as low-coverage areas and will receive focused attention and reinforcement in subsequent training through an active imputation mechanism.

[0070] This invention generates multiple sets of high-quality candidate trajectories in parallel using a first planning model, significantly reducing the search time for online planning. Simultaneously, the backoff mechanism of a second planning model ensures planning success rate and system stability in challenging scenarios. By constructing a closed-loop system of "generation-verification-backoff-learning," the system can automatically identify performance limitations of the first planning model (i.e., low coverage areas) and use the second planning model to generate high-value samples for incremental training, enabling the system to continuously learn and improve, effectively enhancing planning capabilities in boundary and challenging scenarios. The output of the first planning model is always under the monitoring of a deterministic constraint verification and safe backoff framework, ensuring the safety and executability of the final output trajectory. Utilizing the characteristics of generative models, diverse candidate trajectories covering different solution spaces can be generated for the same planning request, effectively addressing the multi-solution problem of redundant robotic arms and reducing the risk of planning failure due to "pattern collapse."

[0071] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0072] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A robot motion planning method, characterized in that, include: Step S10: The system receives a planning request from the upper-layer application. The planning request is a fetching task and includes corresponding planning information. Step S20: The received planning request is encoded into planning condition c by the condition encoding module, and K latent variables z are sampled from a preset probability distribution; Step S30: Using a generative programming network, output K sets of candidate trajectories in parallel based on the planning condition c and K latent variables z; Step S40: Perform constraint verification on the K groups of candidate trajectories and filter out candidate trajectories that fail the constraint verification; Step S50: Score and rank the candidate trajectories that have passed the constraint verification; Step S60: Select one or more candidate trajectories with the highest scores as the initial values ​​for local optimization, and perform a finite number of optimizations, time parameterization, or collision margin correction; Step S70: If the number of candidate trajectories that pass the constraint verification is insufficient or local optimization fails, the planner is triggered to back off, and the failure conditions, back-off results and final execution results are written into the active sampling pool. Step S80: Output the final executable trajectory and record the planning time, pass rate, failure type, optimization residual and safety verification results for incremental training of the generative planning network.

2. The robot motion planning method according to claim 1, characterized in that, The planning information includes: current joint status, target pose / path, obstacles, constraint parameters, and task preferences.

3. The robot motion planning method according to claim 1, characterized in that, The constraint verification includes filtering candidate trajectories that violate end-effector accuracy, joint limit, collision, self-collision, velocity / acceleration, and smoothness constraints.

4. The robot motion planning method according to claim 1, characterized in that, The scoring and ranking includes: ranking multi-objective scores according to error, collision margin, trajectory length, energy consumption, distance from the current attitude, and task preference.

5. A robot motion planning system, characterized in that, include: Module M10: The system receives a planning request from an upper-layer application. The planning request is a fetching task and contains corresponding planning information. Module M20: Encodes the received planning request into planning condition c through the condition coding module, and samples K latent variables z from a preset probability distribution; Module M30: Outputs K sets of candidate trajectories in parallel based on the planning condition c and K latent variables z through a generative programming network; Module M40: Performs constraint verification on K groups of candidate trajectories and filters out candidate trajectories that fail the constraint verification; Module M50: Scores and sorts the candidate trajectories that have passed the constraint verification; Module M60: Select one or more candidate trajectories with the highest scores as initial values ​​for local optimization, and perform a finite number of optimizations, time parameterization, or collision margin correction; Module M70: If the number of candidate trajectories that pass the constraint verification is insufficient or local optimization fails, the planner is triggered to back off, and the failure conditions, back-off results and final execution results are written into the active sampling pool. Module M80: Outputs the final executable trajectory and records the planning time, pass rate, failure type, optimization residual, and safety verification results for incremental training of the generative planning network.

6. The robot motion planning system according to claim 5, characterized in that, The planning information includes: current joint status, target pose / path, obstacles, constraint parameters, and task preferences.

7. The robot motion planning system according to claim 5, characterized in that, The constraint verification includes filtering candidate trajectories that violate end-effector accuracy, joint limit, collision, self-collision, velocity / acceleration, and smoothness constraints.

8. The robot motion planning system according to claim 5, characterized in that, The scoring and ranking includes: ranking multi-objective scores according to error, collision margin, trajectory length, energy consumption, distance from the current attitude, and task preference.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the robot motion planning method according to any one of claims 1 to 4.

10. A robot, characterized in that, Includes the robot motion planning system as described in any one of claims 5-8.

Images