Robot motion planning method and device based on velocity field, equipment and medium

Abstract

The application discloses a robot motion planning method and device based on a velocity field, equipment and a medium. The method comprises the following steps: generating a candidate point set according to an environment grid of a motion area, and determining a high-density sample set of each sensitive area in the environment grid according to the candidate point set; generating a first sample set, and training an initial velocity field estimation network by using the first sample set; generating a second sample set according to the environment grid and a local obstacle transformation condition, and training the initial velocity field estimation network according to the second sample set; generating a velocity field according to a current motion target of the robot and the trained velocity field estimation network, and determining a discrete trajectory and a relative velocity between each trajectory point in the discrete trajectory according to the velocity field. By using the above technical scheme, the velocity field estimation network can be quickly and accurately established, the network has the perception ability of dynamic changes of the environment, can output a reliable velocity field, and the motion planning efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of motion control technology, and in particular to robot motion planning methods, devices, equipment and media based on velocity fields. Background Technology

[0002] Robot motion planning is the process of adapting to the diverse shapes and dynamic behaviors of obstacles in the working environment to generate a collision-free path for the robot's autonomous operation.

[0003] Among them, field-based motion planning methods can quickly generate motion trajectories. Specifically, this involves constructing a field in the robot's motion environment and using the physical properties or mathematical rules of the field to guide the robot from the initial state to the target state. However, the original field-based motion planning methods require manually constructing a high-quality field in advance and lack an efficient update mechanism.

[0004] To improve field construction efficiency, existing technologies have proposed methods for generating velocity fields using neural networks. However, in real-world scenarios with uneven obstacle sizes, these methods tend to overlook small obstacles and oversample large ones, resulting in significant deviations between the generated field and the real environment. This reduces the reliability of the planning and requires a large number of sampling points for neural network training, leading to a long training cycle. Furthermore, when new obstacles appear or the scene changes, the entire network needs to be retrained, making it difficult to achieve rapid iterative updates and adapt to the dynamic changes in the environment perceived in real time during robot movement. Summary of the Invention

[0005] This invention provides a robot motion planning method, device, equipment, and medium based on velocity field, which can quickly and accurately establish a velocity field estimation network. The generated velocity field estimation network has the ability to perceive dynamic changes in the environment and can output a reliable velocity field, thereby effectively improving motion planning efficiency.

[0006] According to one aspect of the present invention, a robot motion planning method based on a velocity field is provided, comprising:

[0007] Based on the environmental grid of the motion area, a set of candidate points is generated, and based on the set of candidate points, a high-density sample set of each sensitive area in the environmental grid is determined;

[0008] Based on the candidate point set and the high-density sample set of each sensitive region, a first sample set is generated, and an initial velocity field estimation network is trained using the first sample set.

[0009] Whenever the motion region completes a local obstacle transformation, a second sample set is generated based on the environmental grid and the local obstacle transformation, and the initial velocity field estimation network is trained based on the second sample set to update the incremental learning network parameters in the initial velocity field estimation network.

[0010] Based on the robot's current motion target and the trained velocity field estimation network, a velocity field is generated, and a discrete trajectory and the relative velocities between each trajectory point in the discrete trajectory are determined based on the velocity field, so as to perform motion control on the robot based on the discrete trajectory and the relative velocities between each trajectory point.

[0011] According to another aspect of the present invention, a robot motion planning device based on a velocity field is provided, comprising:

[0012] The point sampling module is used to generate a set of candidate points based on the environmental grid of the motion area, and to determine a high-density sample set for each sensitive area in the environmental grid based on the set of candidate points.

[0013] The first training module is used to generate a first sample set based on the candidate point set and the high-density sample set of each sensitive region, and to train an initial velocity field estimation network using the first sample set.

[0014] The second training module is used to generate a second sample set based on the environment grid and the local obstacle transformation whenever the motion area completes a local obstacle transformation, and to train the initial velocity field estimation network based on the second sample set to update the incremental learning network parameters in the initial velocity field estimation network.

[0015] The trajectory planning module is used to generate a velocity field based on the robot's current motion target and the trained velocity field estimation network, and to determine a discrete trajectory and the relative velocity between each trajectory point in the discrete trajectory based on the velocity field, so as to perform motion control on the robot based on the discrete trajectory and the relative velocity between each trajectory point.

[0016] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising:

[0017] At least one processor; and

[0018] A memory communicatively connected to the at least one processor; wherein,

[0019] The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to perform the velocity field-based robot motion planning method according to any embodiment of the present invention.

[0020] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to execute and implement the robot motion planning method based on velocity field according to any embodiment of the present invention.

[0021] The technical solution of this invention involves generating a candidate point set based on the environmental grid of the motion region, determining a high-density sample set for each sensitive region in the environmental grid based on the candidate point set, generating a first sample set based on the candidate point set and the high-density sample sets of each sensitive region, and training an initial velocity field estimation network using the first sample set. Whenever the motion region completes a local obstacle transformation, a second sample set is generated based on the environmental grid and the local obstacle transformation, and the initial velocity field estimation network is trained using the second sample set to update the incremental learning network parameters in the initial velocity field estimation network. Based on the robot's current motion target and the trained velocity field estimation network, a velocity field is generated, and a discrete trajectory and various parameters within the discrete trajectory are determined based on the velocity field. The relative velocity between trajectory points, which is used to control the robot's motion based on discrete trajectories and the relative velocities between trajectory points, can solve the problems of unbalanced sampling, low training efficiency, and slow field updates in traditional neural network-based motion planning methods. It significantly reduces the number of sampling points while ensuring that the sampling density in sensitive areas meets training requirements. An accurate velocity field estimation network can be trained with fewer sample points. By updating only some network parameters and using incremental learning, the training speed of the velocity field estimation network can be accelerated. Furthermore, the incremental learning network can achieve rapid field updates for constantly changing motion areas with obstacles. Based on the rapidly updated high-precision velocity field and bidirectional trajectory search, motion planning efficiency can be improved, ensuring collision-free and optimal trajectories.

[0022] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of a robot motion planning method based on a velocity field according to Embodiment 1 of the present invention;

[0025] Figure 2 This is a comparison diagram of sampling point effects provided by an embodiment of the present invention;

[0026] Figure 3 This is a schematic diagram of the structure of a velocity field estimation network according to an embodiment of the present invention;

[0027] Figure 4 This is a flowchart of another robot motion planning method based on a velocity field according to Embodiment 2 of the present invention;

[0028] Figure 5 This is a two-dimensional velocity field thermogram provided by an embodiment of the present invention;

[0029] Figure 6 This is a schematic diagram of a robot motion planning device based on a velocity field according to Embodiment 3 of the present invention;

[0030] Figure 7 This is a schematic diagram of the structure of an electronic device that implements the robot motion planning method based on velocity field according to an embodiment of the present invention. Detailed Implementation

[0031] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0032] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0033] Example 1

[0034] Figure 1This is a flowchart of a robot motion planning method based on a velocity field, provided in Embodiment 1 of the present invention. This embodiment is applicable to situations where precise and rapid motion planning of a robot is performed within a motion region. The method can be executed by a robot motion planning device based on a velocity field. This device can be implemented in hardware and / or software and is generally configured in a computer or processor used for robot control. Figure 1 As shown, the method includes:

[0035] S110. Generate a set of candidate points based on the environmental grid of the motion area, and determine the high-density sample set of each sensitive area in the environmental grid based on the set of candidate points.

[0036] Optionally, an environmental mesh is a mesh structure formed by discretizing the three-dimensional space of the robot's motion area. It can be used to accurately represent the distribution of obstacles in the environment. For example, in an environmental mesh, an obstacle area can be marked as 1, and an obstacle-free area can be marked as 0. An environmental mesh can be constructed by combining simultaneous localization and mapping technology with environmental data collected by multiple sensors and then using three-dimensional data processing tools.

[0037] Optionally, based on the idea of ​​hierarchical dynamic sampling, coarse sampling and fine sampling can be performed sequentially within the motion region. Specifically, coarse sampling is first performed within the motion region to obtain a set of candidate points and identify sensitive points around obstacles. A low-density shell structure is then established around all obstacles using the sensitive points. Finally, the sensitive points are finely expanded based on Gaussian perturbation to generate a high-density sample set.

[0038] The advantage of this setup is that it automatically adjusts the sampling density by utilizing the size, shape, and other characteristics of obstacles, thereby solving the problem of uneven obstacle sampling caused by traditional uniform sampling. The sampling points are closer to the edge of the obstacles, which can significantly reduce the number of sampling points without sacrificing the quality of the field generated by the velocity field estimation network.

[0039] Optionally, based on a preset candidate point density, a set of candidate points can be uniformly and randomly generated in the environment grid. The set of candidate points can be used to initially cover the entire movement area. Each candidate point in the set is represented by a combination of point position coordinates and field value pairs. The field value pairs include the distance between the candidate point and the starting point of the movement, as well as the distance between the candidate point and the nearest obstacle.

[0040] Optionally, the sensitive area can refer to a specific area defined by taking the continuous obstacle area in the environmental grid as the core and combining it with a preset sensitive distance range. The sampling points in this area have a significant impact on the quality of the velocity field generation, and high-density sampling is required to ensure the accuracy of the obstacle surrounding field model.

[0041] Optionally, a high-density sample set refers to a sample set formed by expanding the number of sensitive points in a sensitive area through Gaussian perturbation and then filtering it. Its sampling density is adaptively adjusted according to the size of the sensitive area and the preset point density. The high-density sample set can accurately capture the edge features of obstacles.

[0042] Specifically, generating a candidate point set based on the environmental grid of the motion region, and determining a high-density sample set for each sensitive region in the environmental grid based on the candidate point set, may include:

[0043] Based on the preset candidate point density and the environmental grid, multiple candidate points are uniformly generated to form a candidate point set;

[0044] Calculate the shortest Euclidean distance between each candidate point and the obstacle, and when the shortest Euclidean distance of the target candidate point is within the preset sensitive distance range, the target candidate point is determined as a sensitive point;

[0045] In the environmental grid, at least one sensitive region is determined, and based on the size of the sensitive region and the preset sensitive region point density, the sensitive points in the sensitive region are subjected to at least one round of Gaussian perturbation to generate sensitive candidate points of the sensitive region.

[0046] Once the Gaussian perturbation for the current round is completed, the currently generated sensitive candidate points are filtered according to the sensitive distance range, and a high-density sample set of the sensitive region is generated based on the sensitive points in the sensitive region and the sensitive candidate points retained after filtering.

[0047] Optionally, a uniform random sampling algorithm can be used to generate multiple three-dimensional coordinate points within the effective space of the environmental grid. Each three-dimensional coordinate point can be represented by x. i This means that all generated coordinate points form a candidate point set, which can be represented as X. cand X cand ={x i This ensures that candidate points initially cover the entire motion area, providing comprehensive initial data for subsequent sensitive point identification.

[0048] Optionally, each candidate point in the candidate point set is traversed, and the shortest Euclidean distance from each candidate point to the nearest surrounding obstacle is calculated using the Euclidean distance calculation formula. The preset sensitive distance range can be represented as [d_min, d_max], for example, d_min=0.1m, d_max=0.5m. The sensitive distance range can be adjusted according to the robot size and motion safety threshold. When the shortest Euclidean distance of the target candidate point is within this range, it is determined as a sensitive point, and all sensitive points form a sensitive point set.

[0049] Optionally, the set of sensitive points can be represented as X. sensitive ={xi ∈X cand |d_min<D(x) i )<d_max}, where D(x) i ) represents the sensitive point x i The shortest Euclidean distance can be understood as each sensitive point in the sensitive point set being a candidate point in the candidate point set whose shortest Euclidean distance falls within the sensitive distance range.

[0050] Optionally, in the environmental grid, a continuous obstacle region consisting of one or more adjacent obstacles can be identified, and the maximum value d_max of a preset sensitive distance range can be extended outward based on the boundary of the continuous obstacle region to form a three-dimensional spatial region surrounding the continuous obstacle region, which is the sensitive region.

[0051] It is understandable that the size of the sensitive area within the movement area varies, and the number of sensitive points within each sensitive area also varies. If the number of sensitive points is increased according to a preset magnification factor, it may result in a large number of sampling points in a small area or an insufficient number of sampling points in a large area. Therefore, this invention creatively proposes to calculate the number of target sampling points required for the sensitive area based on the size of the sensitive area and the preset sensitive area point density. For example, the sensitive area point density can be set to include an average of 20 points per grid. If the sensitive area occupies 100 grids, then the number of target sampling points in the sensitive area can be 2000 points, thereby ensuring sufficient sampling density around the obstacle.

[0052] Optionally, Gaussian perturbation can refer to adding a mean of 0 and a variance of δ to the sensitive points. 2 Gaussian noise is used to generate new candidate points. By using Gaussian perturbation, the number of sampling points in the sensitive region can be rapidly expanded. The sensitive candidate points generated by Gaussian perturbation can be represented as x', where x' = x sens +N(0,δ 2 ), where x sens For the set of sensitive points X sensitive Any sensitive point in N(0,δ) 2 ) indicates sensitivity to point x sens Add a mean of 0 and a variance of δ 2 Gaussian noise.

[0053] Optionally, after each round of perturbation, the total number of current sensitive points and sensitive candidate points in the sensitive area is counted. If the target number of sampling points is not reached, the Gaussian perturbation is repeated until the quantity requirement is met or the preset maximum number of perturbation rounds is reached.

[0054] Optionally, after each round of Gaussian perturbation, the shortest Euclidean distance to the nearest obstacle is recalculated for all generated sensitive candidate points. Sensitive candidate points whose shortest Euclidean distance is still within the preset sensitive distance range are selected, and invalid points that are outside the sensitive area or too close or too far from the obstacle are removed. The original sensitive points in the sensitive area are merged with the selected sensitive candidate points to form a high-density sample set for the sensitive area. This sample set can accurately cover the key areas around the obstacle, providing high-quality and balanced sampling data for subsequent network training.

[0055] Figure 2 This is a comparison chart of the effects of one optional sampling point, such as... Figure 2 As shown, the left side represents the high-density sample set generated by the hierarchical dynamic sampling method described in this invention, while the right side represents the sample set generated by multi-point sampling in the prior art. Figure 2 A combination of multiple cubes can be considered as obstacles, and points on or around the obstacles are sampling points. Figure 2 As can be seen from the left image, all sampling points in the high-density sample set collected by this invention are close to the obstacle edge. Even with the superposition of some candidate points from the candidate point set, the overall number of sampling points is still much smaller than the number of sampling points generated by the existing technology on the right. The existing technology, in order to ensure accurate identification of the area around the obstacle by the subsequently trained model, needs to increase the overall sampling density to ensure that the number of sampling points around the obstacle is sufficient for training. Figure 2 As can be seen from the image on the right, in the sample set generated by the existing technology through multi-point sampling, the sampling points are spread out in a certain range around the obstacle, rather than combined with the distribution of the obstacle, and the sampling points are distributed relatively densely.

[0056] Optionally, since the present invention uses a hierarchical dynamic sampling method, the preset point density is relatively small when generating candidate points, and the number of points in the candidate point set is relatively small. During the fine sampling process, the points in the sensitive area can be expanded.

[0057] S120. Generate the first sample set based on the candidate point set and the high-density sample set of each sensitive region, and use the first sample set to train the initial velocity field estimation network.

[0058] Optionally, the velocity field estimation network is a deep learning network built on a physically constrained neural network. The velocity field estimation network proposed in this invention includes three modules: an encoder, a velocity field generation network, and an incremental learning network. By adding an incremental learning network to the velocity field estimation network, this invention can specifically handle scenarios with local environmental changes.

[0059] Optionally, the first sample set can be used for full parameter training of the initial velocity field estimation network. The first sample set includes multiple sample points, each of which is represented by a combination of the sample point's location coordinates and field value pairs. In the first sample set, each sample point is associated with its corresponding standard velocity field value, which is a velocity parameter used to describe the location of the sample point. The standard velocity field value can be pre-calculated based on existing calculation methods.

[0060] Specifically, based on the candidate point set and the high-density sample set of each sensitive region, a first sample set is generated, and an initial velocity field estimation network is trained using the first sample set, including:

[0061] Obtain the union of the candidate point set and each high-density sample set, calculate the standard velocity field value for each sample point in the union, and associate each sample point in the union with the standard velocity field value to generate the first sample set; wherein, the sample point is composed of a sample point position and a field value pair;

[0062] The velocity field estimation network is trained multiple times using the first sample set. After each training round, the prediction loss for the current round is calculated based on the standard velocity field value of the target sample point used in the current training round and the predicted velocity field value of the target sample point output by the velocity field estimation network.

[0063] Based on the predicted loss of the current round, calculate the network update parameters, and update the encoder parameters, velocity field generation network parameters, and incremental learning network parameters in the velocity field estimation network according to the network update parameters of the current round.

[0064] Optionally, the candidate point set and the high-density sample set of all sensitive areas are merged, and duplicate coordinate points are removed to obtain a merged sample point set. Then, for each merged sample point, based on a physical prior method, and combining the three-dimensional coordinates, start coordinates, end coordinates, and obstacle distribution information in the environmental grid, the standard velocity field value S of the sample point is calculated. gt The standard velocity field value can include the optimal arrival time from the starting point to the point, and each sample point is associated with the corresponding standard velocity field value S. gt The first sample set is composed of all data pairs. This sample set takes into account both uniform sampling of the global region and high-density sampling of key regions, which can effectively improve the network training efficiency.

[0065] Optionally, the first sample set can be pre-divided into multiple first sample subsets. In each round of training, a first sample subset is randomly selected as the training data for the current round. Each sample point in the first sample subset used in the current round is the target sample point.

[0066] Figure 3 This is a schematic diagram of an alternative velocity field estimation network. Figure 3 As shown, the velocity field estimation network of the present invention includes an encoder, a velocity field generation network, and an incremental learning network. The encoder is used to encode the model input and send the encoding result to the velocity field generation network. The velocity field generation network is used to generate intermediate prediction results of velocity field values ​​based on the encoding results sent by the encoder. The incremental learning network is used to adjust the intermediate prediction results of velocity field values ​​based on the new field sampling points when the field changes as reported by the robot, and outputs the adjusted predicted velocity field values.

[0067] Optionally, the encoder in the velocity field estimation network can encode each target sample point, and then input the encoding result into the velocity field generation network in the velocity field estimation network to obtain the intermediate prediction result of the velocity field value of each target sample point output by the velocity field generation network. During the network training process in step S120, the obstacles in the motion area are not adjusted. Therefore, when the field changes received by the incremental learning network in this step, the new field sampling point can be defaulted to 0. Finally, the incremental learning network outputs the final predicted velocity field value S for the target sample point based on the intermediate prediction result of the velocity field value sent by the velocity field generation network. pred .

[0068] Optionally, based on the standard velocity field value of the target sample point used in the current training round and the predicted velocity field value of the target sample point output by the velocity field estimation network, the prediction loss of the current round is calculated. Then, the gradient of the prediction loss with respect to all parameters in the velocity field estimation network is calculated through the backpropagation algorithm. Based on the network update parameters of the previous round and the gradient of the prediction loss with respect to the parameters calculated in the current round, the network update parameters of the current round are generated. Based on the network update parameters of the current round, the encoder parameters, velocity field generation network parameters, and incremental learning network parameters in the velocity field estimation network are updated respectively.

[0069] Optionally, when the number of training rounds meets the preset number of rounds or the prediction loss meets the convergence condition, the training ends, the final converged network parameters are saved, and the initial velocity field estimation network is obtained.

[0070] S130. Whenever the motion area completes a local obstacle transformation, a second sample set is generated based on the environmental grid and the local obstacle transformation. The initial velocity field estimation network is then trained based on the second sample set to update the incremental learning network parameters in the initial velocity field estimation network.

[0071] Optionally, after completing step S120 and initially generating the velocity field estimation network, obstacle transformation can be performed on local areas within the motion area multiple times, while keeping the obstacles in other areas unchanged. The completion of local obstacle transformation within the motion area can be confirmed based on the real-time received indication signal or the signal detection results within the motion area.

[0072] Optionally, the second sample set may refer to the incremental dataset generated by resampling based on the updated local environment mesh after each local obstacle transformation, which is used for parameter update training of the incremental learning network.

[0073] Whenever the motion region completes a local obstacle transformation, a second sample set is generated based on the environmental mesh and the local obstacle transformation. The initial velocity field estimation network is then trained using this second sample set to update the incremental learning network parameters within the initial velocity field estimation network. This process may include:

[0074] Whenever the movement area completes a local obstacle change, the environmental mesh is locally updated based on the local obstacle change.

[0075] In the currently updated local environment mesh, determine the incremental candidate point set and the incremental high-density sample set, and generate a second sample set based on the incremental candidate point set and the incremental high-density sample set;

[0076] The encoder parameters and velocity field generation network parameters in the initial velocity field estimation network are fixed, and the initial velocity field estimation network is trained in multiple rounds using the second sample set;

[0077] After each training round, the incremental learning network parameters in the initial velocity field estimation network are updated based on the network update parameters of the current round until the termination condition is met, at which point the trained velocity field estimation network is obtained.

[0078] Optionally, after completing a local obstacle transformation in the movement area, the robot can acquire environmental data of the movement area in real time through sensors, compare and analyze it with historical environmental data, and use point cloud registration or mesh difference detection algorithms to identify local obstacle transformations in the environment. The mesh is updated only in the area of ​​the environmental mesh where the obstacle transformation has occurred and the surrounding extended area, for example, by 0.5m, to ensure coverage of the affected area and keep the mesh data of other areas unchanged.

[0079] Optionally, in the currently updated local environment mesh, the hierarchical dynamic sampling process in step S110 is repeated to generate an incremental candidate point set and an incremental high-density sample set that contain only the local area. The incremental candidate point set is generated in the same way as the candidate point set in step S110, and the incremental high-density sample set is generated in the same way as the high-density sample set in step S110, except that the area range of the generated sampling points is different.

[0080] Optionally, determining the incremental candidate point set and the incremental high-density sample set within the currently updated local environment mesh may include:

[0081] Based on the preset candidate point density, multiple incremental candidate points are uniformly generated in the local environment grid to form an incremental candidate point set.

[0082] Calculate the shortest Euclidean distance between each incremental candidate point and the obstacle, and when the shortest Euclidean distance of the target incremental candidate point is within the preset sensitive distance range, the target incremental candidate point is determined as an incremental sensitive point.

[0083] In the local environment grid, at least one sensitive region is determined, and based on the size of the sensitive region and the preset sensitive region point density, the incremental sensitive points in the sensitive region are subjected to at least one round of Gaussian perturbation to generate incremental sensitive candidate points in the sensitive region.

[0084] Once the Gaussian perturbation for the current round is completed, the currently generated incremental sensitive candidate points are filtered according to the sensitive distance range. Based on the incremental sensitive points in the sensitive area and the incremental sensitive candidate points retained after filtering, an incremental high-density sample set of the sensitive area is generated.

[0085] Optionally, the incremental candidate point set and the incremental high-density sample set are merged, and duplicate coordinate points are removed to obtain a merged sample point set. Then, for each merged sample point, the standard velocity field value of the sample point is calculated based on the physical prior method, and each sample point is associated with the corresponding standard velocity field value. All data pairs form a second sample set.

[0086] Optionally, after generating the second sample set, load the already trained initial velocity field estimation network, fix all parameters of the encoder and velocity field generation network in the network, and the encoder and velocity field generation network do not participate in gradient updates, only allowing parameter updates of the incremental learning network.

[0087] Optionally, the second sample set is input into the velocity field estimation network. The encoder and velocity field generation network of the network perform feature encoding and preliminary field value prediction on the input data based on fixed parameters. The incremental learning network corrects the preliminary prediction results and outputs the final predicted velocity field value. Based on the standard velocity field value and the predicted velocity field value, the prediction loss is calculated. Then, the gradient of the prediction loss with respect to the parameters of the incremental learning network is calculated through the backpropagation algorithm. Based on the gradient of the parameters of the incremental learning network, the parameters of the incremental learning network are updated.

[0088] Optionally, each time the incremental learning network parameters are updated, a local obstacle transformation is performed on another region within the motion area. Then, based on the local region of this transformation, the incremental learning network parameters are updated again until a preset number of incremental training rounds are completed or the loss value of the incremental learning network converges. At this point, training stops, and the updated incremental learning network parameters are combined with the fixed encoder and velocity field generation network parameters to obtain the trained velocity field estimation network.

[0089] Optionally, by fixing the parameters of the encoder and velocity field generation network, only the relevant parameters of the incremental learning network are updated. This can quickly adapt to local changes in the environment, achieve rapid iteration of the field, avoid the inefficiency caused by retraining the entire network, and significantly improve the field update efficiency.

[0090] S140. Based on the robot's current motion target and the trained velocity field estimation network, generate a velocity field, and determine the discrete trajectory and the relative velocity between each trajectory point in the discrete trajectory based on the velocity field, so as to perform motion control on the robot based on the discrete trajectory and the relative velocity between each trajectory point.

[0091] Optionally, the velocity field estimation network can output a field estimation matrix based on the input start point, end point and environmental information, and then generate a velocity field. The velocity field is a three-dimensional spatial field model that represents the optimal arrival time from the start point to each point in the motion area in the form of contour lines. The vertical direction of the contour lines is the collision-free optimal driving path, and its field value relationship can be converted into the relative speed between each trajectory point.

[0092] Optionally, a discrete trajectory can refer to a set of ordered trajectory points generated based on the velocity field gradient. Based on the discrete trajectory and the relative velocities between the trajectory points, the robot control system can generate a strategy and perform precise control of the robot's motion based on the control strategy.

[0093] Optionally, the target starting position, target ending position, and change point cloud information of the robot within the motion area are input into the updated velocity field estimation network. The velocity field estimation network outputs a field estimation matrix of fixed size, which is then visualized to form the velocity field.

[0094] Optionally, based on the symmetry of the velocity field, trajectory points are selected from both the target starting position and the target ending position, along the direction of the steepest change in the velocity field gradient, ultimately forming a discrete trajectory connecting the starting point and the ending point.

[0095] Optionally, based on the field value relationship of each trajectory point in the velocity field, the relative velocity between each trajectory point can be determined. The discrete trajectory and relative velocity information are sent to the robot control system through a specified communication mechanism. After the robot control system solves the problem, it determines the control parameters of the robot motion mechanism and performs motion control on the robot to achieve precise movement of the robot along the discrete trajectory.

[0096] Optionally, the applicable scenarios of the present invention include, but are not limited to, robotic arms, drones, legged robots, autonomous driving, and other scenarios.

[0097] The technical solution of this invention involves generating a candidate point set based on the environmental grid of the motion region, determining a high-density sample set for each sensitive region in the environmental grid based on the candidate point set, generating a first sample set based on the candidate point set and the high-density sample sets of each sensitive region, and training an initial velocity field estimation network using the first sample set. Whenever the motion region completes a local obstacle transformation, a second sample set is generated based on the environmental grid and the local obstacle transformation, and the initial velocity field estimation network is trained using the second sample set to update the incremental learning network parameters in the initial velocity field estimation network. Based on the robot's current motion target and the trained velocity field estimation network, a velocity field is generated, and a discrete trajectory and various parameters within the discrete trajectory are determined based on the velocity field. The relative velocity between trajectory points, which is used to control the robot's motion based on discrete trajectories and the relative velocities between trajectory points, can solve the problems of unbalanced sampling, low training efficiency, and slow field updates in traditional neural network-based motion planning methods. It significantly reduces the number of sampling points while ensuring that the sampling density in sensitive areas meets training requirements. An accurate velocity field estimation network can be trained with fewer sample points. By updating only some network parameters and using incremental learning, the training speed of the velocity field estimation network can be accelerated. Furthermore, the incremental learning network can achieve rapid field updates for constantly changing motion areas with obstacles. Based on the rapidly updated high-precision velocity field and bidirectional trajectory search, motion planning efficiency can be improved, ensuring collision-free and optimal trajectories.

[0098] Example 2

[0099] Figure 4 This is a flowchart illustrating a robot motion planning method based on a velocity field, as provided in Embodiment 2 of the present invention. This embodiment, based on the above embodiments, specifically describes the robot motion planning method based on a velocity field. Figure 4 As shown, the method includes:

[0100] S210. Based on the preset candidate point density and environmental grid, generate multiple candidate points uniformly to form a candidate point set.

[0101] S220. Calculate the shortest Euclidean distance between each candidate point and the obstacle, and when the shortest Euclidean distance of the target candidate point is within the preset sensitive distance range, determine the target candidate point as a sensitive point.

[0102] S230. In the environmental grid, at least one sensitive region is determined, and based on the size of the sensitive region and the preset sensitive region point density, the sensitive points in the sensitive region are subjected to at least one round of Gaussian perturbation to generate sensitive candidate points of the sensitive region.

[0103] In the environmental grid, at least one sensitive region is determined, and based on the size of the sensitive region and a preset sensitive region point density, the sensitive points within the sensitive region are subjected to at least one round of Gaussian perturbation to generate sensitive candidate points for the sensitive region. This may include:

[0104] In the environmental grid, a continuous obstacle region consisting of at least one obstacle is identified, and a sensitive region is determined based on the maximum value of the sensitive distance range and the boundary of the continuous obstacle region;

[0105] The number of target sampling points within the sensitive area is determined based on the size of the sensitive area and the point density within the sensitive area.

[0106] At least one round of Gaussian perturbation is performed on each sensitive point within the sensitive region to generate sensitive candidate points within the sensitive region. The Gaussian perturbation ends when the number of sampling points in the high-density sample set meets the target number of sampling points.

[0107] Optionally, in the environmental grid, identifying a continuous obstacle region consisting of at least one obstacle, and determining a sensitive region based on the maximum value of the sensitive distance range and the boundary of the continuous obstacle region, may include:

[0108] The algorithm iterates through all obstacle marker cells in the environment mesh and uses a connected component analysis algorithm to identify interconnected obstacle marker cell clusters; each cluster corresponds to a continuous obstacle region.

[0109] Using the outermost grid cell of the continuous obstacle region as the boundary, the maximum value of the preset sensitive distance range is extended outward along each coordinate axis of the three-dimensional space to form a closed three-dimensional region, and the closed three-dimensional region is determined as the sensitive region of the corresponding continuous obstacle region.

[0110] Optionally, by counting the number of environmental grid cells contained in the sensitive area and combining the volume of each grid cell, the total volume of the sensitive area can be calculated. Then, based on the preset point density of the sensitive area and the total volume of the sensitive area, the minimum number of sampling points required for the sensitive area can be calculated to ensure that the sampling data can fully characterize the spatial features around the obstacle.

[0111] The advantage of this setup is that it can accurately match the range of sensitive areas with the distribution of obstacles. The Gaussian perturbation rounds are adaptively adjusted according to the number of target sampling points, which not only ensures the sampling density of key areas around obstacles, but also avoids data redundancy caused by invalid sampling, providing targeted and high-quality sampling data for subsequent network training.

[0112] S240. After determining that the Gaussian perturbation of the current round has been completed, the currently generated sensitive candidate points are filtered according to the sensitive distance range, and a high-density sample set of the sensitive area is generated based on the sensitive points in the sensitive area and the sensitive candidate points retained after filtering.

[0113] S250. Obtain the candidate point set and the union of each high-density sample set. Calculate the standard velocity field value for each sample point in the union and associate each sample point in the union with the standard velocity field value to generate the first sample set.

[0114] The sample point consists of the sample point location and the field value pair. The field value pair of the sample point may include the distance between the sample point and the starting point of the motion and the distance between the sample point and the nearest obstacle.

[0115] S260. The velocity field estimation network is trained in multiple rounds using the first sample set. After each round of training, the prediction loss for the current round is calculated based on the standard velocity field value of the target sample point used in the current training round and the predicted velocity field value of the target sample point output by the velocity field estimation network.

[0116] The calculation of the prediction loss for the current training round, based on the standard velocity field value of the target sample points used in the current training round and the predicted velocity field value of the target sample points output by the velocity field estimation network, can include:

[0117] Calculate the confidence level of the target sample point based on the standard velocity field value and the predicted velocity field value of the target sample point;

[0118] The prediction loss for the current round is calculated based on the confidence level, standard velocity field value, and predicted velocity field value of each target sample point.

[0119] Optionally, the confidence level c of the target sample point j can be calculated using the following formula. j :

[0120] ;

[0121] Among them, S gt (x j S represents the standard velocity field value of the target sample point j. pred (x j ) represents the predicted velocity field value of the target sample point j.

[0122] Optionally, the prediction loss L for the target sample point in the current round can be calculated using the following formula. FI :

[0123] ;

[0124] Where N is the number of sample points input in this round of training, and c j S represents the confidence level of target sample point j in this round of training. gt (x j S represents the standard velocity field value of the target sample point j. pred (x j () represents the predicted velocity field value of the target sample point j. l This serves as a marker for calculating standard L2 distance. This can represent the standard L2 distance between the calculated standard velocity field value and the predicted velocity field value of the target sample point j. The standard L2 distance is also known as the mean square error. γ is an adjustable focusing parameter that can be preset.

[0125] S270. Based on the predicted loss of the current round, calculate the network update parameters, and based on the network update parameters of the current round, update the encoder parameters, velocity field generation network parameters, and incremental learning network parameters in the velocity field estimation network respectively.

[0126] S280. Whenever the movement area completes a local obstacle transformation, the environmental mesh is locally updated based on the local obstacle transformation.

[0127] S290. In the currently updated local environment mesh, determine the incremental candidate point set and the incremental high-density sample set, and generate a second sample set based on the incremental candidate point set and the incremental high-density sample set.

[0128] S2100, fix the encoder parameters in the initial velocity field estimation network and the velocity field generation network parameters, and use the second sample set to train the initial velocity field estimation network in multiple rounds.

[0129] S2110. After each training round, update the incremental learning network parameters in the initial velocity field estimation network according to the network update parameters of the current round until the termination condition is met, and then obtain the trained velocity field estimation network.

[0130] S2120. Based on the robot's current motion target and the trained velocity field estimation network, generate a velocity field, and determine the discrete trajectory and the relative velocity between each trajectory point in the discrete trajectory based on the velocity field, so as to perform motion control on the robot based on the discrete trajectory and the relative velocity between each trajectory point.

[0131] This process, which involves generating a velocity field based on the robot's current motion target and the trained velocity field estimation network, and determining a discrete trajectory and the relative velocities between points on that trajectory based on the velocity field, may include:

[0132] Based on the robot's current motion target and current position, determine the target's starting point position and target's ending point position at the current moment, and determine the changed point cloud information based on the point cloud data collected by the robot at the current moment and historical point cloud data.

[0133] The target starting position, target ending position, and changed point cloud information are input into the trained velocity field estimation network. Based on the field estimation matrix output by the velocity field estimation network, a velocity field is generated.

[0134] Based on the velocity field, trajectory points are determined simultaneously from the target starting position and the target ending position, and the relative velocities between the trajectory points are determined based on the coordinates of each trajectory point and the velocity field.

[0135] Optionally, when the robot starts, the preset starting position of the task can be determined as the target starting position based on the current motion target. During the robot's movement, the robot's current position can be determined as the target starting position, and the preset ending position of the task can be determined as the target ending position based on the robot's current motion target.

[0136] Optionally, during the robot's movement, real-time point cloud data acquired by the robot's sensors at the current moment can be collected, compared with historically stored point cloud data, and the changed point cloud information can be obtained through a difference extraction algorithm. After converting the changed point cloud information into a coordinate format that matches the environmental grid, it can be input into the velocity field estimation network along with the target starting position and the target ending position.

[0137] Optionally, each element in the field estimation matrix corresponds to the field value at a location in the environmental grid. By performing three-dimensional visualization processing on the field estimation matrix, a velocity field presented in the form of contour lines can be obtained. The numerical value of the contour lines represents the optimal arrival time from the starting point to that location, and the field value corresponding to the obstacle area is infinite.

[0138] Figure 5 This is an optional two-dimensional plane velocity field thermogram. For example... Figure 5As shown, the numbers at the bottom and left of the velocity field on the left are used to identify different coordinate positions in the two-dimensional plane. The right side indicates that the velocity field value 0-1 is represented by different colors or pixel values. The higher the field value, the higher the optimal arrival priority from that position to the starting point or ending point. The closed, continuous, thin curves in the figure are the contour lines of the velocity field. Each line corresponds to the same field value. The thicker, irregular lines in the figure represent obstacles in the environmental grid.

[0139] Optionally, the density of path points in the discrete trajectory can be preset. Based on the velocity field and the preset path point density, the motion trajectory can be searched in both forward and reverse directions simultaneously. Forward search can refer to starting from the target starting position and iteratively generating a sequence of forward trajectory points. Reverse search can refer to starting from the target ending position and iteratively generating a sequence of reverse trajectory points. When the distance between the last point of the forward trajectory and the last point of the reverse trajectory is less than the path point density, the two trajectory point sequences are spliced ​​together to obtain a complete discrete trajectory, which is the optimal path for collision-free motion.

[0140] Optional, discrete trajectory τ * It can be represented as Where τ1 can be the target starting position at the current calculation time, τ end The target endpoint location can be used as the basis for the forward trajectory generation logic, which can be represented as follows: , where τ i For any point in the discrete trajectory, τ i+1 For τ i The next trajectory point, the positive trajectory is generated sequentially from τ1 towards the target endpoint, β is the pre-set path point density, Q represents the velocity field, Q(x s x goal ) represents the region in the velocity field from the starting position of the target to the ending position of the target, ▽τ i This indicates that it needs to be based on τ i Determine τ i The point where the change is steepest between the target endpoint and the next trajectory point τ is selected as the next trajectory point. i+1 The location.

[0141] Optionally, the reverse trajectory generation logic can be represented as follows: , where τ i For any point in the discrete trajectory, τ i-1 For τ i The previous trajectory point, the reverse trajectory is from τ end The trajectory points are generated sequentially towards the target starting position, where β is the pre-defined path point density, Q represents the velocity field, and Q(x) goal x s) represents the region in the velocity field from the target's endpoint position to the target's starting position, ▽τ i This indicates that it needs to be based on τ i Determine τ i The point where the change is steepest between the target starting point and the target starting point is taken as the previous trajectory point τ. i-1 The location.

[0142] Optionally, the contour lines in the velocity field represent the optimal time to reach each trajectory point, and the relative velocity between each trajectory point can be obtained based on the optimal time of each trajectory point.

[0143] The technical solution of this invention involves generating a candidate point set based on the environmental grid of the motion region, determining a high-density sample set for each sensitive region in the environmental grid based on the candidate point set, generating a first sample set based on the candidate point set and the high-density sample sets of each sensitive region, and training an initial velocity field estimation network using the first sample set. Whenever the motion region completes a local obstacle transformation, a second sample set is generated based on the environmental grid and the local obstacle transformation, and the initial velocity field estimation network is trained using the second sample set to update the incremental learning network parameters in the initial velocity field estimation network. Based on the robot's current motion target and the trained velocity field estimation network, a velocity field is generated, and a discrete trajectory and various parameters within the discrete trajectory are determined based on the velocity field. The relative velocity between trajectory points, which is used to control the robot's motion based on discrete trajectories and the relative velocities between trajectory points, can solve the problems of unbalanced sampling, low training efficiency, and slow field updates in traditional neural network-based motion planning methods. It significantly reduces the number of sampling points while ensuring that the sampling density in sensitive areas meets training requirements. An accurate velocity field estimation network can be trained with fewer sample points. By updating only some network parameters and using incremental learning, the training speed of the velocity field estimation network can be accelerated. Furthermore, the incremental learning network can achieve rapid field updates for constantly changing motion areas with obstacles. Based on the rapidly updated high-precision velocity field and bidirectional trajectory search, motion planning efficiency can be improved, ensuring collision-free and optimal trajectories.

[0144] Example 3

[0145] Figure 6 This is a schematic diagram of a robot motion planning device based on a velocity field, provided in Embodiment 3 of the present invention. Figure 6 As shown, the device includes: a point sampling module 310, a first training module 320, a second training module 330, and a trajectory planning module 340.

[0146] The point sampling module 310 is used to generate a set of candidate points based on the environmental grid of the motion area, and to determine a high-density sample set for each sensitive area in the environmental grid based on the set of candidate points.

[0147] The first training module 320 is used to generate a first sample set based on the candidate point set and the high-density sample set of each sensitive region, and to train an initial velocity field estimation network using the first sample set.

[0148] The second training module 330 is used to generate a second sample set based on the environment grid and the local obstacle transformation whenever the motion area completes a local obstacle transformation, and to train the initial velocity field estimation network based on the second sample set to update the incremental learning network parameters in the initial velocity field estimation network.

[0149] The trajectory planning module 340 is used to generate a velocity field based on the robot's current motion target and the trained velocity field estimation network, and to determine a discrete trajectory and the relative velocity between each trajectory point in the discrete trajectory based on the velocity field, so as to perform motion control on the robot based on the discrete trajectory and the relative velocity between each trajectory point.

[0150] The technical solution of this invention involves generating a candidate point set based on the environmental grid of the motion region, determining a high-density sample set for each sensitive region in the environmental grid based on the candidate point set, generating a first sample set based on the candidate point set and the high-density sample sets of each sensitive region, and training an initial velocity field estimation network using the first sample set. Whenever the motion region completes a local obstacle transformation, a second sample set is generated based on the environmental grid and the local obstacle transformation, and the initial velocity field estimation network is trained using the second sample set to update the incremental learning network parameters in the initial velocity field estimation network. Based on the robot's current motion target and the trained velocity field estimation network, a velocity field is generated, and a discrete trajectory and various parameters within the discrete trajectory are determined based on the velocity field. The relative velocity between trajectory points, which is used to control the robot's motion based on discrete trajectories and the relative velocities between trajectory points, can solve the problems of unbalanced sampling, low training efficiency, and slow field updates in traditional neural network-based motion planning methods. It significantly reduces the number of sampling points while ensuring that the sampling density in sensitive areas meets training requirements. An accurate velocity field estimation network can be trained with fewer sample points. By updating only some network parameters and using incremental learning, the training speed of the velocity field estimation network can be accelerated. Furthermore, the incremental learning network can achieve rapid field updates for constantly changing motion areas with obstacles. Based on the rapidly updated high-precision velocity field and bidirectional trajectory search, motion planning efficiency can be improved, ensuring collision-free and optimal trajectories.

[0151] Based on the above embodiments, the point sampling module 310 may include:

[0152] The candidate point set generation unit is used to uniformly generate multiple candidate points according to the preset candidate point density and the environmental grid, forming a candidate point set;

[0153] The sensitive point selection unit is used to calculate the shortest Euclidean distance between each candidate point and the obstacle, and when the shortest Euclidean distance of the target candidate point is within the preset sensitive distance range, the target candidate point is determined as a sensitive point.

[0154] A sensitive candidate point generation unit is used to determine at least one sensitive region in the environmental grid, and to perform at least one round of Gaussian perturbation on the sensitive points in the sensitive region according to the size of the sensitive region and the preset sensitive region point density, so as to generate sensitive candidate points of the sensitive region.

[0155] The high-density sample set generation unit is used to filter the currently generated sensitive candidate points according to the sensitive distance range after determining that the Gaussian perturbation of the current round has been completed, and to generate a high-density sample set of the sensitive region based on the sensitive points in the sensitive region and the sensitive candidate points retained after filtering.

[0156] Based on the above embodiments, the sensitive candidate point generation unit can be specifically used for:

[0157] In the environmental grid, a continuous obstacle region consisting of at least one obstacle is identified, and a sensitive region is determined based on the maximum value of the sensitive distance range and the boundary of the continuous obstacle region;

[0158] The number of target sampling points within the sensitive area is determined based on the size of the sensitive area and the point density within the sensitive area.

[0159] At least one round of Gaussian perturbation is performed on each sensitive point within the sensitive region to generate sensitive candidate points within the sensitive region. The Gaussian perturbation ends when the number of sampling points in the high-density sample set meets the target number of sampling points.

[0160] Based on the above embodiments, the first training module 320 may include:

[0161] The first sample set generation unit is used to obtain the candidate point set and the union of each high-density sample set, calculate the standard velocity field value for each sample point in the union, and associate each sample point in the union with the standard velocity field value to generate the first sample set; wherein, the sample point is composed of a sample point position and a field value pair.

[0162] The prediction loss calculation unit is used to train the velocity field estimation network for multiple rounds using the first sample set, and after each round of training, it calculates the prediction loss for the current round based on the standard velocity field value of the target sample point used in the current training round and the predicted velocity field value of the target sample point output by the velocity field estimation network.

[0163] The parameter update unit is used to calculate the network update parameters based on the prediction loss of the current round, and update the encoder parameters, velocity field generation network parameters, and incremental learning network parameters in the velocity field estimation network according to the network update parameters of the current round.

[0164] Based on the above embodiments, the prediction loss calculation unit can be specifically used for:

[0165] Calculate the confidence level of the target sample point based on the standard velocity field value and the predicted velocity field value of the target sample point;

[0166] The prediction loss for the current round is calculated based on the confidence level, standard velocity field value, and predicted velocity field value of each target sample point.

[0167] Based on the above embodiments, the second training module 330 can be specifically used for:

[0168] Whenever the movement area completes a local obstacle change, the environmental mesh is locally updated based on the local obstacle change.

[0169] In the currently updated local environment mesh, determine the incremental candidate point set and the incremental high-density sample set, and generate a second sample set based on the incremental candidate point set and the incremental high-density sample set;

[0170] The encoder parameters and velocity field generation network parameters in the initial velocity field estimation network are fixed, and the initial velocity field estimation network is trained in multiple rounds using the second sample set;

[0171] After each training round, the incremental learning network parameters in the initial velocity field estimation network are updated based on the network update parameters of the current round until the termination condition is met, at which point the trained velocity field estimation network is obtained.

[0172] Based on the above embodiments, the trajectory planning module 340 can be specifically used for:

[0173] Based on the robot's current motion target and current position, determine the target's starting point position and target's ending point position at the current moment, and determine the changed point cloud information based on the point cloud data collected by the robot at the current moment and historical point cloud data.

[0174] The target starting position, target ending position, and changed point cloud information are input into the trained velocity field estimation network. Based on the field estimation matrix output by the velocity field estimation network, a velocity field is generated.

[0175] Based on the velocity field, trajectory points are determined simultaneously from the target starting position and the target ending position, and the relative velocities between the trajectory points are determined based on the coordinates of each trajectory point and the velocity field.

[0176] The robot motion planning device based on velocity field provided in the embodiments of the present invention can execute the robot motion planning method based on velocity field provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0177] Example 4

[0178] Figure 7 A schematic diagram of an electronic device 10, which can be used to implement embodiments of the present invention, is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0179] like Figure 7 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0180] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0181] Processor 11 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, central processing unit (CPU), graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as the velocity field-based robot motion planning method described in the embodiments of the present invention. That is:

[0182] Based on the environmental grid of the motion area, a set of candidate points is generated, and based on the set of candidate points, a high-density sample set of each sensitive area in the environmental grid is determined;

[0183] Based on the candidate point set and the high-density sample set of each sensitive region, a first sample set is generated, and an initial velocity field estimation network is trained using the first sample set.

[0184] Whenever the motion region completes a local obstacle transformation, a second sample set is generated based on the environmental grid and the local obstacle transformation, and the initial velocity field estimation network is trained based on the second sample set to update the incremental learning network parameters in the initial velocity field estimation network.

[0185] Based on the robot's current motion target and the trained velocity field estimation network, a velocity field is generated, and a discrete trajectory and the relative velocities between each trajectory point in the discrete trajectory are determined based on the velocity field, so as to perform motion control on the robot based on the discrete trajectory and the relative velocities between each trajectory point.

[0186] In some embodiments, the velocity field-based robot motion planning method can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the velocity field-based robot motion planning method described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to execute the velocity field-based robot motion planning method by any other suitable means (e.g., by means of firmware).

[0187] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0188] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0189] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0190] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0191] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0192] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0193] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0194] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A robot motion planning method based on velocity field, characterized in that, include: Based on the environmental grid of the motion area, a set of candidate points is generated, and based on the set of candidate points, a high-density sample set of each sensitive area in the environmental grid is determined; Based on the candidate point set and the high-density sample set of each sensitive region, a first sample set is generated, and an initial velocity field estimation network is trained using the first sample set; wherein, the first sample set includes multiple sample points, and each sample point is associated with a corresponding standard velocity field value. Whenever the motion region completes a local obstacle transformation, a second sample set is generated based on the environmental grid and the local obstacle transformation. The initial velocity field estimation network is then trained based on the second sample set to update the incremental learning network parameters in the initial velocity field estimation network. The second sample set is an incremental dataset generated by resampling based on the updated local environmental grid after each local obstacle transformation. Based on the robot's current motion target and the trained velocity field estimation network, a velocity field is generated, and a discrete trajectory and the relative velocities between each trajectory point in the discrete trajectory are determined based on the velocity field, so as to perform motion control on the robot based on the discrete trajectory and the relative velocities between each trajectory point. Specifically, a candidate point set is generated based on the environmental grid of the motion region, and a high-density sample set for each sensitive region in the environmental grid is determined based on the candidate point set, including: Based on the preset candidate point density and the environmental grid, multiple candidate points are uniformly generated to form a candidate point set; Calculate the shortest Euclidean distance between each candidate point and the obstacle, and when the shortest Euclidean distance of the target candidate point is within the preset sensitive distance range, the target candidate point is determined as a sensitive point; In the environmental grid, at least one sensitive region is determined, and based on the size of the sensitive region and the preset sensitive region point density, the sensitive points in the sensitive region are subjected to at least one round of Gaussian perturbation to generate sensitive candidate points of the sensitive region. Once the Gaussian perturbation for the current round is completed, the currently generated sensitive candidate points are filtered according to the sensitive distance range, and a high-density sample set of the sensitive region is generated based on the sensitive points in the sensitive region and the sensitive candidate points retained after filtering.

2. The method according to claim 1, characterized in that, In the environmental grid, at least one sensitive region is identified, and based on the size of the sensitive region and a preset sensitive region point density, the sensitive points within the sensitive region are subjected to at least one round of Gaussian perturbation to generate candidate sensitive points for the sensitive region, including: In the environmental grid, a continuous obstacle region consisting of at least one obstacle is identified, and a sensitive region is determined based on the maximum value of the sensitive distance range and the boundary of the continuous obstacle region; The number of target sampling points within the sensitive area is determined based on the size of the sensitive area and the point density within the sensitive area. At least one round of Gaussian perturbation is performed on each sensitive point within the sensitive region to generate sensitive candidate points within the sensitive region. The Gaussian perturbation ends when the number of sampling points in the high-density sample set meets the target number of sampling points.

3. The method according to claim 1, characterized in that, Based on the candidate point set and the high-density sample set of each sensitive region, a first sample set is generated, and an initial velocity field estimation network is trained using the first sample set, including: Obtain the union of the candidate point set and each high-density sample set, calculate the standard velocity field value for each sample point in the union, and associate each sample point in the union with the standard velocity field value to generate the first sample set; wherein, the sample point is composed of a sample point position and a field value pair; The velocity field estimation network is trained multiple times using the first sample set. After each training round, the prediction loss for the current round is calculated based on the standard velocity field value of the target sample point used in the current training round and the predicted velocity field value of the target sample point output by the velocity field estimation network. Based on the predicted loss of the current round, calculate the network update parameters, and update the encoder parameters, velocity field generation network parameters, and incremental learning network parameters in the velocity field estimation network according to the network update parameters of the current round.

4. The method according to claim 3, characterized in that, Based on the standard velocity field values ​​of the target sample points used in the current training round and the predicted velocity field values ​​of the target sample points output by the velocity field estimation network, the prediction loss for the current round is calculated, including: Calculate the confidence level of the target sample point based on the standard velocity field value and the predicted velocity field value of the target sample point; The prediction loss for the current round is calculated based on the confidence level, standard velocity field value, and predicted velocity field value of each target sample point.

5. The method according to claim 1, characterized in that, Whenever the motion region completes a local obstacle transformation, a second sample set is generated based on the environmental mesh and the local obstacle transformation. The initial velocity field estimation network is then trained using this second sample set to update the incremental learning network parameters within the initial velocity field estimation network, including: Whenever the movement area completes a local obstacle change, the environmental mesh is locally updated based on the local obstacle change. In the currently updated local environment mesh, determine the incremental candidate point set and the incremental high-density sample set, and generate a second sample set based on the incremental candidate point set and the incremental high-density sample set; The encoder parameters and velocity field generation network parameters in the initial velocity field estimation network are fixed, and the initial velocity field estimation network is trained in multiple rounds using the second sample set; After each training round, the incremental learning network parameters in the initial velocity field estimation network are updated based on the network update parameters of the current round until the termination condition is met, at which point the trained velocity field estimation network is obtained.

6. The method according to claim 1, characterized in that, Based on the robot's current motion target and the trained velocity field estimation network, a velocity field is generated, and a discrete trajectory and the relative velocities between points on the discrete trajectory are determined based on the velocity field, including: Based on the robot's current motion target and current position, determine the target's starting point position and target's ending point position at the current moment, and determine the changed point cloud information based on the point cloud data collected by the robot at the current moment and historical point cloud data. The target starting position, target ending position, and changed point cloud information are input into the trained velocity field estimation network. Based on the field estimation matrix output by the velocity field estimation network, a velocity field is generated. Based on the velocity field, trajectory points are determined simultaneously from the target starting position and the target ending position, and the relative velocities between the trajectory points are determined based on the coordinates of each trajectory point and the velocity field.

7. A robot motion planning device based on a velocity field, characterized in that, include: The point sampling module is used to generate a set of candidate points based on the environmental grid of the motion area, and to determine a high-density sample set for each sensitive area in the environmental grid based on the set of candidate points. The first training module is used to generate a first sample set based on the candidate point set and the high-density sample set of each sensitive region, and to train an initial velocity field estimation network using the first sample set; wherein, the first sample set includes multiple sample points, and each sample point is associated with a corresponding standard velocity field value. The second training module is used to generate a second sample set based on the environment grid and the local obstacle transformation whenever the motion region completes a local obstacle transformation, and to train the initial velocity field estimation network based on the second sample set to update the incremental learning network parameters in the initial velocity field estimation network; wherein, the second sample set is an incremental dataset generated by resampling based on the updated local environment grid after each local obstacle transformation; The trajectory planning module is used to generate a velocity field based on the robot's current motion target and the trained velocity field estimation network, and to determine a discrete trajectory and the relative velocity between each trajectory point in the discrete trajectory based on the velocity field, so as to perform motion control on the robot based on the discrete trajectory and the relative velocity between each trajectory point. The point sampling module includes: The candidate point set generation unit is used to uniformly generate multiple candidate points according to the preset candidate point density and the environmental grid, forming a candidate point set; The sensitive point selection unit is used to calculate the shortest Euclidean distance between each candidate point and the obstacle, and when the shortest Euclidean distance of the target candidate point is within the preset sensitive distance range, the target candidate point is determined as a sensitive point. A sensitive candidate point generation unit is used to determine at least one sensitive region in the environmental grid, and to perform at least one round of Gaussian perturbation on the sensitive points in the sensitive region according to the size of the sensitive region and the preset sensitive region point density, so as to generate sensitive candidate points of the sensitive region. The high-density sample set generation unit is used to filter the currently generated sensitive candidate points according to the sensitive distance range after determining that the Gaussian perturbation of the current round has been completed, and to generate a high-density sample set of the sensitive region based on the sensitive points in the sensitive region and the sensitive candidate points retained after filtering.

8. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the robot motion planning method based on the velocity field as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the robot motion planning method based on a velocity field as described in any one of claims 1-6.

Images