Robot multi-constraint path planning method and device based on diffusion model, electronic equipment, storage medium and program product

Abstract

The invention discloses a robot multi-constraint path planning method and device based on a diffusion model, electronic equipment, a storage medium and a program product. The method comprises the following steps: collecting environment data of a multi-source sensor and constructing a map; based on a search strategy, global parameter optimization and local differentiable smoothing, a high-quality path data set meeting reachability, smoothness and multi-class constraints is generated cooperatively; training a diffusion model by using a data set, and learning kinematics, safety and geometric constraints in noise prediction through multi-constraint joint loss; introducing environmental condition guidance during reasoning, and alternately executing constraint gradient guidance and feasible region projection to dynamically correct an intermediate path; and selecting an optimal path from the candidate paths and generating a robot control instruction to realize closed-loop trajectory tracking. According to the method, a dual consistency mechanism combining training constraint and reasoning constraint is provided, the safety, feasibility and real-time performance of path planning are remarkably improved, and the method is suitable for robot autonomous navigation in a complex constraint environment.

Description

TECHNICAL FIELD

[0001] The application relates to the field of robot technology and artificial intelligence technology, and in particular to a robot multi-constraint path planning method and device based on a diffusion model, an electronic device, a storage medium and a program product. BACKGROUND

[0002] Path planning is the core ability of a robot system to achieve autonomous motion and is widely used in industrial production, warehouse logistics, commercial services, automatic driving, forest inspection and special robots. Traditional rule-based graph search methods (such as Hybrid A*) can generate a feasible path that meets kinematic constraints in a static environment, but such methods rely on artificially defined cost functions and heuristic strategies, are highly sensitive to key parameters, are difficult to maintain global optimality in a multi-constraint coupled scene, and have low planning efficiency in a complex environment, making it difficult to meet real-time requirements.

[0003] To improve planning quality, existing methods introduce global optimization or local optimization strategies, such as using meta-heuristic algorithms to search for key parameters or using conjugate gradient, quasi-Newton and other local optimization methods to smooth the path. However, such optimization methods usually require a large number of iterative calculations, are sensitive to initial solutions, and still consume high computational resources in high-dimensional, multi-constraint environments, making it difficult to meet real-time deployment requirements.

[0004] In recent years, machine learning-based methods have been gradually used for path planning tasks, including reinforcement learning, deep learning and end-to-end learning methods. Such methods can learn the environment structure and decision-making rules from data to a certain extent, improving the adaptability of the model to complex scenes. However, reinforcement learning is extremely sensitive to reward design, has poor strategy stability and high training cost; deep neural networks can simulate complex spatial structures, but generally lack strict physical and constraint feasibility guarantees.

[0005] With the development of generative modeling technology, end-to-end generative path planning methods have become a new trend. Generative models can learn the latent planning distribution from a large number of example paths, directly obtaining smooth, continuous and efficient paths through a single generation process, thereby significantly reducing the computational overhead caused by traditional search and optimization. As the most representative generative model, diffusion models have become an important research direction in generative path planning due to their strong distribution modeling ability, multi-modal characteristics and stable training mechanism. Diffusion models generate feasible paths from random noise through a reverse step-by-step "denoising" approach and have achieved remarkable results in trajectory prediction and behavior generation tasks.

[0006] However, despite the potential of diffusion models in trajectory generation, they still have significant limitations in complex real-world path planning scenarios. On the one hand, the path distribution learned by diffusion models often relies on the quality of the training data, and it is difficult to maintain global optimality under multiple constraints such as collision safety, speed and acceleration boundaries, terrain passability, etc. The generated path is prone to local feasibility but overall suboptimal. On the other hand, as a probabilistic generative model, the generation process of diffusion models lacks built-in constraint compliance, and it is easy to produce small but critical constraint violations, such as slight overlap between the path and obstacles, small turning beyond the kinematic limits of the robot, etc., which makes it difficult to meet the safety and reliability requirements of actual robot systems. These limitations hinder the further development of diffusion models from "generating feasible paths" to "generating reliable and optimal paths".

[0007] In addition, existing technologies related to diffusion models also have different degrees of limitations:

[0008] (1) The planning methods based on diffusion models (such as Diffuser, TrajDiff, MotionDiff) still have the following limitations:

[0009] · Mostly used for unconstrained or weakly constrained trajectory prediction, lacking unified modeling of kinematic, geometric structure and collision constraints;

[0010] · The training process focuses on trajectory fitting or reward optimization, and the reasoning stage lacks dynamic constraint correction, leading to output results that are prone to failure in complex scenarios;

[0011] · Mostly using one-time sampling or post-global correction, unable to guide and project intermediate results in real time during denoising;

[0012] · Difficult to handle high-complexity planning tasks with multiple objectives, multiple scales and multiple constraint couplings.

[0013] (2) Traditional optimization-based planning methods (such as MPC, gradient optimization) have the following limitations:

[0014] · Relies on explicit mathematical models, limited adaptability to environmental changes, non-convex or sparse obstacle structures;

[0015] · Prone to local optima, highly sensitive to initial paths or heuristics;

[0016] · Lack of multi-modal path generation capability, unable to generalize path distribution through offline learning.

[0017] (3) Path planning methods based on deep reinforcement learning (DRL) face the following problems:

[0018] · Sensitive to reward functions, training is unstable;

[0019] • The generated strategy lacks constraint compliance guarantee;

[0020] • In complex structure scenarios, it is easy to produce uninterpretable or even dangerous behavior;

[0021] • High training cost and weak migration limit engineering application.

[0022] In summary, path planning technology is evolving from rule-based search and optimization to learning-based generative planning. Although diffusion models have significant advantages in generation efficiency and trajectory smoothness, they still cannot meet the actual system requirements in terms of optimality and safety due to the lack of high-quality training data, the lack of integration of multi-constraint information, and the lack of dynamic feasibility guarantee in the reasoning process.

[0023] To solve the above problems, the present application provides a robot multi-constraint path planning method based on diffusion model, device, electronic equipment, storage medium and program product, by constructing high-quality path data set, embedding multi-class constraint loss in training stage, and performing guiding and feasible region projection operation based on environment information in reasoning stage, realizing the unified improvement of path generation quality and multi-constraint compliance, thereby significantly improving the shortcomings of the prior art in complex scenarios. SUMMARY

[0024] I. Invention purpose

[0025] The existing generative path planning method generally has the problems of low path quality, insufficient response to complex environmental constraints, difficulty in ensuring kinematic feasibility, difficulty in effectively avoiding collision risk, and difficulty in balancing real-time performance in a multi-constraint environment, which limits its application in mobile robots, autonomous driving, complex terrain navigation and other scenarios.

[0026] The purpose of the present application is to provide a multi-constraint path planning method based on diffusion model and related device to improve the quality of path generation, adapt to complex constraint conditions, realize real-time planning ability in dynamic environment, and improve the overall safety and reliability.

[0027] II. Technical solutions (1) Multi-constraint path planning method based on diffusion model

[0028] To achieve the above purpose, the present application provides a multi-constraint path planning method based on diffusion model, which comprises:

[0029] Step 1: Environment data acquisition

[0030] Synchronously collect laser radar point cloud, visual image and inertial measurement unit data, time stamp align and coordinate system unify the multi-source heterogeneous data, and form spatiotemporally consistent multi-modal environment perception data.

[0031] Step two: map construction and data preprocessing

[0032] Point cloud registration and gridding processing are performed on the multi-modal data to generate a passable area cost map, which is then encoded into a normalized standard feature tensor.

[0033] Step three: constructing a high-quality path dataset

[0034] A path sample set for training is obtained, which can be obtained in multiple ways, including simulation planner generated samples, real robot running data, or artificial teaching trajectories. The key parameters of various path generation strategies are optimized and adjusted, and local smoothing, geometric consistency processing, and feasibility checking are performed on the generated path to obtain a high-quality path dataset with global accessibility, local smoothness, and constraint compliance.

[0035] Step four: training the diffusion model based on multi-constraint joint loss

[0036] The diffusion model is trained using the path dataset. During training, a multi-constraint joint loss function is constructed to make the diffusion model consider kinematic constraints, geometric environment constraints, collision risk metrics, physical consistency indicators, and task semantic preferences based on noise prediction targets, to optimize path generation quality and satisfaction of multiple constraints.

[0037] Step five: performing path guidance and constraint projection in the inference stage

[0038] During the step-by-step denoising inference process of the diffusion model, the intermediate path is updated according to the environment information, constraint conditions, and task target, and the path is fine-tuned towards a feasible and low-risk direction. At the same time, the intermediate path is projected to a feasible set that satisfies spatial geometry, kinematic feasibility, or task constraints through constraint projection operations, to maintain the real-time feasibility and safety of the path at each inference step. The guidance and projection operations can be alternately performed to continuously modify the intermediate generated path.

[0039] Step six: selecting the optimal path and issuing execution

[0040] After multiple iterations to generate multiple candidate paths that satisfy the constraints, the system selects the optimal path according to the task preference (such as the shortest path, the smoothest path), and converts it into line speed and angular velocity control instructions for the robot bottom layer. Through the combination of real-time feedback from the odometer and IMU, a closed-loop motion control is formed to drive the robot to accurately and smoothly track the generated path and complete the navigation task.

[0041] (II) Multi-constraint path planning device based on diffusion model

[0042] The application also provides a multi-constraint path planning device based on a diffusion model, comprising:

[0043] Module one: multi-source data acquisition module

[0044] It is used for synchronously collecting multi-source heterogeneous original perception data of a robot, including laser radar point cloud, depth vision image, IMU (inertial measurement unit) and wheeled odometry information.

[0045] Module two: data preprocessing module

[0046] This module performs point cloud processing, map construction and multi-source fusion on original perception data, and finally outputs a standardized multi-channel feature tensor, providing a unified input for downstream planning and decision-making.

[0047] Module two: path data construction module

[0048] It is used for acquiring and constructing the above high-quality path data set, and generating training samples meeting feasibility and constraint compliance through multi-class planning strategies and local optimization means.

[0049] Module three: denoising generation module

[0050] It is used for performing a reverse denoising process of the diffusion model to gradually generate a structured path sequence from an initial random path.

[0051] Module four: conditional guidance module

[0052] It is used for embedding map structure, environment attribute, start and end points, task conditions and other multi-class conditional information in the denoising process, and guiding the path generation direction through feature modulation or conditional fusion.

[0053] Module five: training and constraint alignment module

[0054] It is used for collaborative optimization based on noise prediction loss and multi-constraint joint loss to promote the model to learn the path generation ability consistent with the constraints.

[0055] Module six: path correction and output module

[0056] It is used for performing guided update based on a target function or constraint condition in the inference stage, and performing constraint projection operation, and finally outputting a path planning result meeting the multi-constraint conditions.

[0057] Module seven: robot control and execution module

[0058] It is used for converting the planned path into robot chassis control instructions. Its core is real-time calculation and closed-loop feedback control based on the differential kinematics model: the linear speed and angular speed are calculated according to the expected trajectory, and the dynamic adjustment is continuously detected and tracked to realize high-precision trajectory tracking and smooth running in complex scenes.

[0059] (Three) Electronic device

[0060] The present application also provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable by the processor, the program implementing steps one to four of the above method when executed.

[0061] (Four) Non-transitory computer readable storage medium

[0062] The present application provides a non-transitory computer readable storage medium having computer instructions stored thereon, which, when executed by a processor, cause the processor to implement the above method.

[0063] (Five) Computer program product

[0064] The present application provides a computer program product comprising computer program instructions, which, when executed on a computer device, cause the computer to execute the above path planning method.

[0065] Third, the innovation and benefits of the present application

[0066] Compared with the prior art, the present application proposes a new generative planning framework that deeply integrates diffusion models with multi-constraint path planning, with the following innovations and technical effects:

[0067] (1) A two-stage guarantee mechanism for multi-constraint consistency is proposed. The present application introduces a multi-constraint joint loss in the diffusion model training stage and a guidance and constraint projection mechanism based on environmental information in the inference stage, achieving two-stage constraint consistency guarantee for multiple geometric constraints, kinematic constraints and safety constraints. This mechanism is different from the traditional model which only handles constraints during training or after planning, significantly improving the stability and reliability of path generation.

[0068] (2) A "denoising-guiding-projection" closed-loop generation structure is proposed. The present application first embeds guiding gradients and feasible region projection into the diffusion inference step, so that the intermediate path generated by each denoising can be corrected in time, forming a continuous closed-loop control chain. This structure avoids the problem of uncontrollable cumulative error in the generation process of traditional diffusion models and can stably generate paths that meet multiple constraints in complex environments.

[0069] (3) A high-quality path dataset generated by search strategy, optimization strategy and heuristic parameter tuning is constructed. The present application constructs a high-quality path sample distribution for training through adaptive optimization of planner parameters, local differentiable smoothing processing and path geometric consistency verification, so that the diffusion model can learn a distribution form closer to the real executable path, improving the model generation ability from the source.

[0070] (4) High scalability and task adaptability. The multi-constraint joint loss and projection mechanism of the application have scalability, and new constraint types can be flexibly added according to different task requirements, covering complex environments, dynamic obstacles, multi-robot cooperation and other scenes.

[0071] (5) Realize high real-time performance and engineering deployability. The reasoning structure of the application reduces the invalid sampling process through constraint guidance, and guarantees the path feasibility through the projection mechanism, thereby reducing the post-processing overhead, and realizing efficient real-time path generation under limited computing power conditions.

[0072] In summary, the application proposes a new technical system in the field of generative path planning, which can effectively solve the shortcomings of the prior art in complex constraint environments, and has significant novelty and substantial progress. BRIEF DESCRIPTION OF DRAWINGS

[0073] In order to more clearly illustrate the technical solutions in the disclosure or related art, the following will briefly introduce the drawings needed to be used in the embodiments or related art description. Obviously, the drawings in the following description are only embodiments of the disclosure, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0074] Figure 1 An application scenario schematic diagram of the robot multi-constraint path planning method based on the diffusion model provided by the embodiments of the application is shown in the figure.

[0075] Figure 2 A flowchart of the robot multi-constraint path planning method based on the diffusion model provided by the embodiments of the application is shown in the figure.

[0076] Figure 3 A structural schematic diagram of the robot multi-constraint path planning device based on the diffusion model provided by the embodiments of the application is shown in the figure.

[0077] Figure 4 An electronic device hardware structure schematic diagram of the embodiments of the application is shown in the figure. DETAILED DESCRIPTION

[0078] In order to make the purpose, technical scheme and advantages of the disclosure clearer, the principles and spirits of the disclosure will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are only given to enable those skilled in the art to better understand and implement the disclosure, and do not limit the scope of the disclosure in any way. On the contrary, these embodiments are provided to make the disclosure more thorough and complete, and to fully convey the scope of the disclosure to those skilled in the art.

[0079] In this document, it is to be understood that the number of elements in any figure of the drawings is presented for example only and not limitation, and any naming is made merely for the purpose of differentiating between different components and not limitation.

[0080] It should be noted that the technical terms or scientific terms used in the embodiments of the present disclosure are used in the general sense unless otherwise defined.

[0081] The terms should be understood in their ordinary sense to those of ordinary skill in the art to which the present disclosure belongs. The terms "first", "second" and similar words used in the embodiments of the present disclosure do not represent any order, number or importance, but are only used to distinguish different components. As described in the background, existing path planning algorithms face many challenges in multi-constraint complex scenarios. In the robot navigation task, generating a high-quality path that meets both kinematic and dynamic constraints and can respond to environmental changes in real time is the key to improving overall planning performance. Existing methods are usually based on sampling or optimization, and there is a significant trade-off between path feasibility, constraint satisfaction, and computational efficiency. When the environmental constraints are complex, the path dimension is high, or there is a multi-objective optimization requirement, traditional methods often struggle to generate globally optimal and locally smooth path results within a limited time.

[0082] In the multi-constraint path planning problem, there may be coupling or even conflict between various constraint conditions (such as path length, terrain structure, and motion smoothness). The existing algorithms lack effective multi-constraint collaborative optimization mechanisms, resulting in poor performance in spatial continuity, motion rationality, and real-time feasibility of the generated path. In addition, most generative path planning methods lack explicit alignment mechanisms for multi-constraint conditions during the training phase, making it difficult to ensure the satisfaction of complex constraints by the generated path. In the inference process, there is also a lack of effective online guidance and real-time correction capability, further limiting its practical application effect and system robustness in high-dynamic, multi-constraint scenarios.

[0083] The article "a" or "an" before an element does not exclude the presence of more than one of that element.

[0084] The principles and spirits of the present disclosure will be explained in detail below with reference to several representative embodiments of the present disclosure.

[0085] As described in the background, existing path planning algorithms face many challenges in multi-constraint complex scenarios. In the robot navigation task, generating a high-quality path that meets both kinematic and dynamic constraints and can respond to environmental changes in real time is the key to improving overall planning performance. Existing methods are usually based on search or optimization, and there is a significant trade-off between path feasibility, constraint satisfaction, and computational efficiency. When the environmental constraints are complex, the path dimension is high, or there is a multi-objective optimization requirement, traditional methods often struggle to generate globally optimal and locally smooth path results within a limited time.

[0086] In multi-constraint path planning problems, various constraints (such as path length, terrain structure, and motion smoothness) may be coupled or even conflict. Existing algorithms lack effective multi-constraint collaborative optimization mechanisms, resulting in insufficient performance of generated paths in terms of spatial continuity, motion rationality, and real-time feasibility. Furthermore, most generative path planning methods heavily rely on high-quality training data and lack explicit alignment mechanisms for multiple constraints during the training phase, making it difficult to guarantee the satisfaction of complex constraints by the generated paths. Simultaneously, they lack effective online guidance and real-time correction capabilities during inference. These limitations further restrict their practical application effectiveness and system robustness in highly dynamic, multi-constraint scenarios.

[0087] To overcome the limitations of existing technologies, this invention proposes a novel multi-constraint path planning method for robots based on a diffusion model. This method first integrates multi-level path optimization strategies during the path dataset generation process to construct a high-quality path dataset that combines global optimality and local smoothness. Based on this, model training is performed by applying forward noise to the path data and utilizing a UNet network to perform multi-step back-diffusion denoising, enabling the model to learn the ability to recover complete and smooth paths from noisy paths. During training, a collaborative optimization mechanism of constraint alignment loss and noise prediction loss is introduced simultaneously to enhance the generated path's ability to comply with complex constraints. In the model inference stage, a conditional guidance mechanism embeds map information and multi-source conditions such as start and end points into the denoising process, and the FILM mechanism is used to achieve feature fusion of conditional information and the UNet network. In each denoising iteration, gradient optimization and constraint projection operations are combined to dynamically correct the generated path, ultimately outputting a high-quality path planning result that satisfies multiple constraints. This method has significant application value and promising prospects in fields such as autonomous robot navigation, unmanned driving, and personalized navigation services.

[0088] After introducing the basic principles of the present invention, various non-limiting embodiments of the present invention will be described in detail below.

[0089] refer to Figure 1 This is a schematic diagram illustrating an application scenario of a robot multi-constraint path planning system based on a diffusion model in an exemplary embodiment of the present invention. The system includes a robot terminal 101, a server 102, and a data storage system 103. The components are connected via wired or wireless communication networks to realize the uploading of environmental perception data, the deployment of path planning tasks, and the storage and management of execution results.

[0090] The robot terminal 101 is a mobile robot with autonomous navigation capabilities, integrating sensors such as LiDAR, depth camera, odometry, and inertial measurement unit (IMU). The LiDAR is used to collect environmental point cloud information, the depth camera is used to acquire the 3D structure of near-field obstacles, and the odometry and IMU are used to estimate the robot's pose and motion state. Through multi-source sensor information fusion, the robot terminal can construct an environmental map and execute the path planning method proposed in this invention, achieving real-time trajectory generation and optimization in dynamic environments.

[0091] Server 102 serves as the system's control hub, deployed at edge nodes or in the cloud. It processes environmental data uploaded by the robot, performs global path planning and multi-robot collaborative scheduling, and dynamically adjusts planning parameters according to task requirements to improve the overall operating efficiency of the system in complex scenarios.

[0092] The data storage system 103 is used to store environment models, path planning results, historical task data and system parameters. It has an environment model library and operation monitoring functions, supports path reuse and rapid task reconstruction, effectively reduces the burden of repetitive calculations in the system and improves response speed.

[0093] The system's components achieve high-speed and reliable data transmission via local area network, industrial Wi-Fi, or 5G network, ensuring the system's communication needs in real-time path planning and multi-machine collaboration. It should be understood that... Figure 1 This system is for illustrative purposes only and can be flexibly configured according to actual application scenarios. It is suitable for autonomous navigation tasks in various constrained environments such as warehousing and logistics, indoor inspection, and intelligent delivery.

[0094] It is important to note that Figure 1 The application scenario shown is merely an example of the present invention, used to help understand the technical principles of the invention. In practical applications, the system described in this invention can be flexibly deployed according to different environmental scales and task requirements, and is suitable for path planning and autonomous navigation scenarios in warehousing logistics, indoor inspection, intelligent delivery, and other confined spaces.

[0095] refer to Figure 2 The aforementioned robot multi-constraint path planning based on a diffusion model includes the following steps:

[0096] Step S201: Environmental Data Acquisition

[0097] During actual operation, the robot terminal utilizes multiple sensors, including LiDAR, depth cameras, inertial measurement units (IMUs), and wheeled odometry, to comprehensively perceive the operating environment and capture real-time information on surrounding obstacles, ground structures, and boundaries. The system performs time synchronization and spatial registration on various types of raw data to construct a multimodal environment dataset that integrates point clouds, depth images, and attitude information. To improve perception accuracy in complex scenarios such as narrow passages, the data acquisition module employs an anomaly suppression strategy based on temporal filtering to effectively correct for instantaneous noise and sensor drift. This module provides complete and accurate environmental information input for subsequent map building and path planning.

[0098] Step S202: Map Building and Data Preprocessing

[0099] First, voxel filtering is performed on the acquired point cloud data to achieve data sparsity, effectively improving system storage efficiency and computing performance. Then, the ICP algorithm or its improved version is used for laser point cloud registration, completing high-precision stitching of multiple frames of point cloud data.

[0100] Based on point cloud processing, the system constructs a local occupancy grid map and converts it into a two-dimensional cost matrix to quantify environmental accessibility. During this process, attitude information from the inertial measurement unit and wheel velocity data are simultaneously fused to reduce cumulative system error.

[0101] The processed local cost map, location information, and other multi-source data are uniformly encoded into a fixed-dimensional multi-channel tensor. Simultaneously, numerical normalization is performed to form standardized input, providing a unified interface for subsequent model training and inference.

[0102] Step S203: Generation of high-quality path dataset

[0103] To meet the requirements of path planning in terms of smoothness, safety, and minimizing the number of reversals, we improved the Hybrid A* algorithm. Specifically, by introducing a safety distance constraint and considering vehicle turning, we optimized the algorithm's cost function:

[0104] f(s i )=g(s i )+h(s i )

[0105] Where f(s) i ) represents the current node s i The total value of the product, g(s) i ) represents the distance from the starting point to the current node s. i The cumulative cost value, h(s) i It uses the traditional Hybrid A* heuristic function.

[0106] The main optimization is the cumulative cost value g(s). i ):

[0107]

[0108] cost(s n-1 ,s n ) indicates from node s n-1 To the next node s n The extended cost function, d(s) i-1 ,s i ) represents the distance between adjacent nodes, δ i The symbol represents the direction of movement, with 1 representing forward and -1 representing backward, but the default is δ. i Greater than 0, w r ,w t These are the turning penalty coefficient and the steering penalty coefficient. i The risk cost represents the current state, based on the robot's position and the distance d from the obstacle. ob Calculations show that d min For safe distance.

[0109] Subsequently, the lemming optimization algorithm is used to optimize key parameters, such as heuristic cost weights and turning penalty coefficients. The lemming optimization algorithm is a novel metaheuristic algorithm inspired by four key group behaviors exhibited by lemmings in nature: long-distance migration, burrowing, foraging in groups, and escaping predators. This algorithm simulates these behaviors mathematically, aiming to efficiently solve complex optimization problems. Its core objective is to guide the search agent to explore and develop the solution space, ultimately finding a globally optimal or near-optimal solution.

[0110] The cost function of robot path planning is typically complex, non-convex, and non-linear, and contains multiple local optima. The lemming optimization algorithm possesses a good balance between global exploration and local exploitation, effectively escaping local optima and searching for better solutions in a broad parameter space. More importantly, it exhibits better convergence performance and faster convergence speed compared to other metaheuristic algorithms.

[0111] Subsequently, the L-BFGS algorithm is employed for local optimization. This algorithm does not require explicit construction or storage of the Hessian matrix; it approximates its inverse by maintaining historical iteration vectors, significantly reducing memory overhead while ensuring superlinear convergence. Combined with the Wolfe conditional line search mechanism, L-BFGS achieves stable and efficient local optimization, making it suitable for high-dimensional non-convex problems.

[0112] The lemming optimization algorithm and the L-BFGS algorithm jointly optimize the same objective function, with the objective function for path P set to the following form:

[0113] J(P)=wlen J len (P)+w cur J cur (P)+w smo J smo (P)+w obs J obs (P)

[0114] Where J(P) represents the path length term, J cur (P) represents the curvature term, J smo (P) represents the smoothness term, J obs (P) represents the obstacle avoidance term. The corresponding coefficient w represents the weight coefficient of each term, and p... i p represents the i-th path point on the path. i =(x i ,y i Let J(P) be expressed as a two-dimensional coordinate vector. Then the total gradient of J(P) is... for:

[0115]

[0116] L-BFGS path update method is α is determined by the L-BFGS line search.

[0117] Step S204: Train the diffusion model to incorporate alignment constraints

[0118] The diffusion model is a powerful generative model whose core idea is to learn to recover the distribution of the target data from pure noisy data through a framework of progressively adding noise (forward process) and progressively removing noise (backward process).

[0119] The forward pass is a fixed, learning-free process. It progressively adds Gaussian noise to a clean path of data A0 through a series of steps t = 1...T.

[0120]

[0121] Where q(A) t |A t-1 ) is a conditional probability distribution, describing the entire process as a Markov chain. Each step is scheduled according to a predefined variance β. t The process was repeated, ultimately resulting in pure noise A that was completely unrecognizable. T .

[0122] This process can directly calculate the noise data A at any time t from T0 using a closed-form solution. t :

[0123]

[0124] Where α t =1-β t ,

[0125] Train the neural network parameters θ to directly predict the noise ε added to the data A0 during the forward pass. θ The loss function is defined as the prediction noise ε. θ Mean square error between the actual noise ε and the true noise ε:

[0126]

[0127] Here, O represents conditional information (such as map images and start and end points). By minimizing this loss, the neural network learns how to denoise any noisy path into a feasible, collision-free path using a reverse process based on environmental information.

[0128] The goal of the reverse process is to transform pure noise A T The process begins by removing noise step by step, ultimately generating a path A0 that conforms to the original data distribution (i.e., reasonable and feasible). Each denoising operation is defined by a learned Gaussian distribution:

[0129]

[0130] Where θ are the neural network parameters. The goal of this network is to predict how to obtain a less noisy path at step t-1, given a noisy path at step t.

[0131] Diffusion models can generate high-quality and diverse trajectory solutions, but purely data-driven training often leads to violations of key constraints, such as reaching the target, avoiding collisions, and following system dynamics. Therefore, a hybrid loss function is introduced to explicitly measure and penalize constraint violations during training.

[0132] First, a collision loss is introduced. To ensure that the generated path maintains a safe distance from obstacles, while maintaining the differentiability of the loss function to support gradient backpropagation, this invention constructs an obstacle avoidance loss based on continuous differentiability.

[0133] V obs (p,o)=max(d safe -d,0) 2

[0134] Specifically, for any point p on the predicted path ξ and an obstacle o, the degree of safety distance violation is defined as follows:

[0135] V obs (p,o)=max(d safe -d,0)2

[0136] Where d = ||po||2, the function violates the boundary condition d = d safe It has a continuous first derivative, which effectively avoids gradient abrupt changes and ensures training stability.

[0137] For the entire discrete trajectory ξ, the total obstacle avoidance loss is defined as the sum of the degrees of violation of all obstacles at all trajectory points:

[0138]

[0139] Where d = ||po||2, the function violates the boundary condition d = d safe The loss function has a continuous first derivative, effectively avoiding abrupt gradient changes and ensuring training stability. It is integrated into the framework of constraint alignment diffusion training and serves as a core component of constraint violation loss. It generates trajectories that meet safety distance requirements through gradient-guided diffusion models.

[0140] Then, a kinematic constraint loss is introduced. To ensure that the generated path conforms to the robot's own kinematic characteristics, while maintaining the differentiability of the loss function to support gradient backpropagation, this invention constructs a curvature constraint loss based on continuous differentiability.

[0141] Specifically, for any three consecutive path points p on the discrete trajectory ξ i-1 ,p i ,p i+1 First, its approximate value at p is calculated. i Discrete curvature κ at i :

[0142]

[0143] Its curvature violation is defined as:

[0144] V curv (p t )=max(|κ t |-κ max ,0) 2

[0145] Among them κ max This is the maximum curvature threshold that the robot can execute. The function is effective when the boundary |κ is violated. t |-κ max It has a continuous first derivative, which effectively avoids gradient abrupt changes and ensures training stability.

[0146] For the entire discrete trajectory ξ, the total kinematic constraint loss is defined as the sum of curvature violations at all intermediate trajectory points:

[0147]

[0148] The loss function It is integrated into the constraint-aware diffusion training framework and serves as a key component of the constraint violation loss. It generates smooth trajectories that meet kinematic feasibility requirements through gradient-guided diffusion models.

[0149] Step S205: Generate a compliant path by combining cost gradient and projection operations.

[0150] First, the gradient guidance method is used to optimize and guide path points to move in the compliant direction during the diffusion model inference process. Its core advantage is that it integrates trajectory generation and optimization in the same diffusion process. It is not a traditional form of post-processing gradient optimization and can avoid the shortcomings of post-processing gradient optimization, such as sensitivity to initial values ​​and local optima.

[0151] The specific process involves first calculating the path prior mean in each denoising step:

[0152]

[0153] Where β t Let α be the variance scheduling value at time step t. t =1-β t , ε represents the prediction noise.

[0154] Subsequently, according to μ t Calculate the guiding gradient g:

[0155]

[0156] in Let λ be the gradient of the cost function with different constraints. i These are the corresponding weighting coefficients.

[0157] Then update the guiding mean.

[0158]

[0159] Finally, sample the next state τ. t-1 :

[0160]

[0161] Where z is standard Gaussian noise, σ t The noise intensity is controlled by the variance of the time step t. The entire process continues until the final path τ0 is generated through iteration.

[0162] Gradient guidance, through the gradient field of the loss function, provides an optimization direction toward a safe and smooth region during the generation process of the diffusion model, which is a powerful soft constraint. However, since it is essentially an iterative denoising based on probability distribution, coupled with gradient locality and confidence issues in complex environments, the final samples still have a certain probability of slightly or even significantly violating key constraints (such as collisions).

[0163] Therefore, introducing a hard constraint mechanism based on the projection operator is crucial. After gradient guidance, a projection operator is employed to ensure that the generated result remains within the feasible set C. Map the current sample x to the nearest feasible point y:

[0164]

[0165] Projection operator Its role is to act as a deterministic corrector, which does not rely on probability or gradient estimation, but performs rigorous geometric feasibility verification and correction.

[0166] To strictly enforce obstacle avoidance constraints, we apply a projection operation to pathpoints that fall into the obstacle region. This operation aims to ensure that for each violation point p, in the collision-free free space... Find its Euclidean nearest neighbor, and thus map it to a condition that satisfies the safety boundary. Target point p′:

[0167]

[0168] Among them, free space Maintain at least a safe distance d from obstacles safe It consists of points.

[0169] Step S206: Select the optimal path and transmit it to the robot control module.

[0170] The diffusion model outputs multiple high-quality paths. These paths satisfy differential kinematic constraints, minimum turning radius limits, and obstacle safety distance requirements, exhibiting high continuity and execution stability. The system can select the optimal path based on parameter preferences, demonstrating the advantages of the diffusion model's multimodal generation. This framework is adaptable to different computing platforms. The final results can be directly transmitted to the control module, enabling the robot to achieve autonomous navigation and efficient movement in complex and multi-constraint environments.

[0171] refer to Figure 3This is a schematic diagram of the trajectory recovery device provided in an exemplary embodiment of the present invention. The device includes a multi-source data acquisition module 310, a data preprocessing module 320, a high-quality path dataset generation module 330, a UNet denoising module 340, a condition guidance module 350, a training and constraint alignment module 360, and a trajectory recovery output module 370.

[0172] The multi-source data acquisition module 310 is configured to control the robot system to acquire environmental and motion state information in real time during task execution, enabling comprehensive perception of its surrounding space and changes in its own position. This module may include various sensors such as LiDAR, depth cameras, inertial measurement units (IMUs), and wheeled odometry to simultaneously acquire environmental point clouds, obstacle depth data, and attitude angular velocity and acceleration information. By synchronizing and aligning the sensors in time, this module ensures consistent data sources and continuous temporal sequence, thus providing reliable environmental input for subsequent path calculation. The implementation of this module enables the robot to stably acquire high-precision environmental information even in complex or poorly lit environments, improving the overall system's perception capabilities and path planning accuracy.

[0173] The data preprocessing module 320 is configured to preprocess and standardize the raw environmental data acquired by multi-source sensors using a feedforward neural network. This module uses a multilayer perceptron to filter sensor noise, normalize values, and regularize dimensions, converting heterogeneous inputs such as LiDAR point clouds, IMU data, and visual information into a unified feature tensor. The processing includes data alignment and channel fusion operations to ensure that sensor data from different sampling frequencies remain consistent in the spatiotemporal dimensions. The standardized environmental representation output by this module has fixed dimensions and a stable numerical range, providing a directly processable and effective input for the subsequent high-quality path dataset generation module 330.

[0174] The high-quality path dataset generation module 330 is configured to generate a high-quality path dataset using a multi-level optimization approach with preprocessed environmental information. This module employs an improved Hybrid A* algorithm as the front-end path generator to better satisfy various complex constraints. Then, a lemming optimization algorithm is introduced for mid-range global tuning of key parameters to obtain paths approaching the global optimum. Finally, a gradient optimization algorithm is used for back-end refinement of the paths, thereby constructing a high-quality path dataset that satisfies multiple constraints. This provides a solid data foundation for the diffusion model to generate high-quality paths.

[0175] The UNet denoising module 340 is configured to directly generate high-quality feasible paths based on environmental conditions. This module employs an encoder-decoder architecture, extracting environmental features through a multi-layer convolutional network. During the encoding stage, conditional information such as start and end point positions and obstacle distribution are encoded into high-dimensional feature representations. In the decoding stage, path details are gradually reconstructed through feature upsampling and skip connections, outputting a smooth trajectory that conforms to kinematic constraints. This design achieves an end-to-end mapping from environmental conditions to a complete path, providing an initial path that basically satisfies the distribution of a high-quality path for the subsequent path correction and output module 370.

[0176] The conditional guidance module 350 is configured as an environment feature extractor based on the ResNet-18 architecture, specifically providing multi-level conditional guidance information for the UNet denoising module 340. This module takes the environmental image captured by the forward-facing camera as input and extracts visual features step-by-step through the convolutional layers of ResNet-18, generating feature maps at different scales. These feature maps precisely encode the spatial distribution of obstacles in the environment, the boundaries of feasible regions, and the relative positions of start and end points. In the feature fusion stage, the module injects the extracted deep visual features into the corresponding layers of the UNet decoder via the FiLM mechanism. Specifically, for each layer of the UNet decoder, the conditional guidance module generates corresponding affine transformation parameters (scaling factor γ and bias term β), and precisely integrates environmental constraint information into the path generation process by linearly modulating the intermediate features of UNet. This design ensures that the generated path maintains kinematic feasibility while strictly avoiding obstacles and accurately connecting start and end points.

[0177] The training and constraint alignment module 360 ​​is configured to simultaneously improve trajectory generation quality and constraint satisfaction rate during model training through a joint optimization strategy. The module constructs a composite loss function comprising denoising matching loss and constraint violation loss. The denoising matching loss uses mean squared error to measure the difference between the predicted and true trajectories, while the constraint violation loss quantifies the deviation of the trajectory from physical constraints such as obstacle distance and curvature limits using a differentiable function. To enhance the model's generalization ability, the module introduces multimodal data augmentation strategies during training, including path local pruning and resampling and obstacle noise injection into the map, effectively expanding the distribution range of the training data. For complex constraint optimization problems, the module adopts a progressive training strategy, initially focusing on trajectory reconstruction accuracy and gradually increasing constraint weights in later stages to achieve a balance between generation quality and safety.

[0178] The path correction and output module 370 is configured to generate safe and compliant feasible paths through a two-stage optimization process. This module first introduces a gradient guidance mechanism during path generation, dynamically adjusting the movement direction of path points by analyzing the matching degree between the trajectory and various constraints. This design directly integrates trajectory optimization into the generation process, overcoming the shortcomings of traditional post-processing methods that are sensitive to the initial path and prone to getting trapped in local optima. After gradient guidance, the module further performs geometric projection correction. This step uses spatial geometric calculations to precisely move any path points still in obstacle-prone areas to the nearest safe position, ensuring that all output paths meet preset safety distance requirements. This projection operation does not rely on probabilistic estimation but provides the final guarantee for path safety through deterministic calculations. The double-optimized trajectory is output in a standardized point sequence format, with each path point containing spatial coordinates and motion state information. The generated path shows significant improvements in geometric continuity, dynamic executability, and global safety, thereby ensuring the smooth operation and reliable execution of the mobile robot in complex environments with multiple constraints.

[0179] The control and execution module 380 is configured to receive the optimal path from the path correction and output module 370 and convert it into motion control commands for the robot's underlying system to achieve actual path tracking and execution. This module calculates the attitude deviation in real time based on the robot's current pose and the target path point, and generates corresponding linear and angular velocity control commands. By integrating real-time feedback data from the encoder, odometry, and inertial measurement unit, the module constructs a closed-loop control system to dynamically compensate for and correct path tracking errors. When the perception system detects sudden environmental changes or unexpected obstacles, the module can autonomously trigger a path replanning process, coordinating with the upper-level planning system to generate alternative paths in real time. This control mechanism effectively ensures the robot's continuous operational capability in complex environments, enabling it to smoothly and accurately execute trajectory tracking tasks while maintaining rapid response capabilities to dynamic scenarios.

[0180] It should be noted that the above functional modules are only representative embodiments of the present invention. In actual applications, they can be flexibly adjusted or combined according to different device types and scenario requirements. All equivalent substitutions or modifications made by those skilled in the art based on the content of the present invention should be covered within the protection scope of this patent.

[0181] Figure 4 This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0182] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0183] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0184] The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0185] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0186] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0187] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0188] The electronic devices described above are used to implement the corresponding map-free relocation methods in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0189] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, the present invention also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform the map-free relocation method as described in any of the above embodiments.

[0190] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0191] The aforementioned non-transitory computer-readable storage media can be any available medium or data storage device that a computer can access, including but not limited to magnetic storage (e.g., floppy disks, hard disks, magnetic tapes, magneto-optical disks (MOs), etc.), optical storage (e.g., CDs, DVDs, BDs, HVDs, etc.), and semiconductor storage (e.g., ROMs, EPROMs, EEPROMs, non-volatile memory (NAND flash), solid-state drives (SSDs)).

[0192] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to perform the map-free relocation method as described in any of the embodiments in the exemplary method section above, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0193] Based on the same inventive concept, corresponding to the map-free relocation method described in any of the above embodiments, the present invention also provides a computer program product, which includes computer program instructions. In some embodiments, the computer program instructions can be executed by one or more processors of a computer to cause the computer and / or the processor to perform the map-free relocation method. Corresponding to the execution entity for each step in each embodiment of the map-free relocation method, the processor executing the corresponding step can belong to the corresponding execution entity.

[0194] The computer program product of the above embodiments is used to cause the computer and / or the processor to execute the map-free relocation method as described in any of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0195] Those skilled in the art will recognize that embodiments of the present invention can be implemented as a system, method, or computer program product. Therefore, the present invention can be specifically implemented as entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, the present invention can also be implemented as a computer program product contained in one or more computer-readable media, which includes computer-readable program code.

[0196] Any combination of one or more computer-readable media may be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example,, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (not exhaustive) of a computer-readable storage medium may include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0197] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0198] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0199] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Python, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as "C" or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network (including a local area network (LAN) or a wide area network (WAN)), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0200] It should be understood that each block of a flowchart and / or block diagram, as well as combinations of blocks in a flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device to produce a machine that, when executed by a computer or other programmable data processing device, creates means for implementing the functions / operations specified in the blocks of the flowchart and / or block diagram.

[0201] These computer program instructions may also be stored in a computer-readable medium that enables a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable medium produce a product comprising an instruction apparatus that implements the functions / operations specified in the boxes of a flowchart and / or block diagram.

[0202] Computer program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, such that the instructions that execute on the computer or other programmable apparatus can provide a process for implementing the functions / operations specified in the boxes of a flowchart and / or block diagram.

[0203] Furthermore, although the operations of the method of the present invention are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all of the operations shown must be performed to achieve the desired result. Rather, the steps depicted in the flowchart may be performed in a different order. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0204] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0205] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0206] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this application (including the claims) is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in the details for the sake of brevity.

[0207] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of the implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this application, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0208] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0209] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.

[0210] While the spirit and principles of the invention have been described with reference to several specific embodiments, it should be understood that the invention is not limited to the disclosed specific embodiments, and the division of aspects does not imply that features in these aspects cannot be combined for benefit; such division is merely for ease of description. The invention is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims. The scope of the appended claims is to be interpreted in the broadest sense, thereby encompassing all such modifications and equivalent structures and functions.

Claims

1. A method for robot multi-constraint path planning based on diffusion model, characterized in that, Comprise: (1) Synchronously collect multi-source perception data from laser radar, visual sensor and inertial measurement unit, and perform time stamp alignment and coordinate system consistent preprocessing; (2) Perform point cloud registration and rasterization processing on the preprocessed multi-source perception data to generate an environment cost map representing the passable area and obstacle distribution; further encode the cost map into a normalized standard feature tensor. (3) Based on search strategy, global parameter optimization and local continuous differentiable smoothing processing, a high-quality path data set is constructed to make the generated path comply with global reachability, local smoothness and multi-class constraint conditions; (4) Use the path data set to train the diffusion model, construct a multi-constraint joint loss, and make the diffusion model learn kinematic constraints, obstacle safety constraints and environmental geometric constraints simultaneously in the noise prediction process; (5) In the step-by-step denoising inference process of the diffusion model, the environment condition is introduced to guide the path, and the intermediate path is sequentially updated based on the gradient of the constraint cost function and projected based on the feasible region to dynamically correct the intermediate generated path and ensure its real-time feasibility and safety; (6) Select the target path from the multiple candidate paths generated by inference, and convert the target path into a motion control instruction executable by the robot to realize closed-loop trajectory tracking and necessary online replanning.

2. The method of claim 1, wherein, The search strategy in step (3) includes: constructing a comprehensive cost function including path cost, environmental risk and behavior penalty term to improve the planning performance of the search in complex environments.

3. The method of claim 1, wherein, The global parameter optimization algorithm in step (3) is any meta-heuristic optimization method that can globally optimize path cost weights and smoothing parameters, and the local smoothing processing algorithm in step (3) is any gradient-based optimization method that can perform continuous differentiable smoothing optimization on the path.

4. The method of claim 1, wherein, The multi-constraint joint loss in step (4) includes at least one of: (1) Obstacle avoidance loss based on path point and obstacle distance; (2) Kinematic constraint loss based on path curvature; (3) Smoothness loss based on path continuity; (4) Dynamics feasibility loss based on physical consistency.

5. The method of claim 1, wherein, The guided update in step (5) includes: calculating the gradient of the constraint cost function based on the path prior mean predicted by the diffusion model, and using the gradient to correct the prior mean to obtain the guided path.

6. The method of claim 1, wherein, The constraint projection in step (5) includes: projecting the path points that violate the collision constraint and kinematic constraint to the nearest point in the feasible region that satisfies the environmental geometric constraint and safety distance constraint.

7. A robot multi-constraint path planning apparatus based on a diffusion model, characterized by, Comprise: (1) A multi-source data acquisition module for synchronously collecting multi-source heterogeneous original perception data of a robot; (2) A data preprocessing module for performing point cloud processing, map construction and format unification on the original perception data; (3) A path data construction module for performing search, global optimization and local smoothing to form a high-quality path data set; (4) A diffusion model training module for training a diffusion model based on a multi-constraint joint loss; (5) An inference guidance and projection module for performing environment condition injection, guided update and constraint projection in the denoising inference stage; (6) a path correction and output module for generating a final compliant path and performing control instruction conversion; (7) a control and execution module for driving the robot to move according to the optimized path and implementing closed-loop tracking control.

8. An electronic device, comprising: A computer program product comprising a processor, a memory, and a computer program stored on the memory and executable by the processor, the program, when executed, implementing the method of any one of claims 1 to 7.

9. A non-transitory computer readable storage medium having stored thereon computer program instructions which, when executed by a processor, implement the method of any one of claims 1 to 7.

10. A computer program product, characterised in that, A computer program product comprising computer program instructions which, when run on a computer or robot terminal, cause it to perform the method of any one of claims 1 to 7.

Images