A robot three-dimensional trajectory planning method based on deep neural network

Abstract

The present invention discloses a robot three-dimensional trajectory planning method based on a deep neural network, comprising: step S1: constructing an environmental map, encoding environmental information in the form of a digital grid; step S2: predicting the trajectory probability distribution in the environmental map through a deep neural network model; step S3: generating non-uniform sampling points based on the trajectory probability distribution, while retaining some global uniform sampling points to form a hybrid sampling strategy; step S4: using an improved RRT* algorithm to search for trajectories, the improved RRT* algorithm optimizes the trajectory cost by dynamically adjusting the step size and rewiring mechanism; step S5: updating environmental information in real time and re-predicting the trajectory probability distribution, triggering local trajectory re-planning to deal with dynamic obstacles. The method of the present invention combines the probability prediction of deep learning with the hybrid sampling strategy, which can significantly improve planning efficiency, shorten the average trajectory length, and reduce computing resource consumption in complex environments.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology and three-dimensional trajectory planning, and relates to a robot three-dimensional trajectory planning method based on a deep neural network. Background Art

[0002] Robot three-dimensional trajectory planning is a key technology in the field of robotics. Its goal is to generate safe, efficient and smooth motion trajectories in complex three-dimensional environments. Traditional robot trajectory planning methods are mainly divided into grid search (such as A), random sampling (such as RRT) and artificial potential field methods. Among them, although sampling algorithms such as RRT can handle high-dimensional space, they have problems such as slow convergence speed and poor initial trajectory quality. In recent years, deep learning technology based on neural networks has been introduced into the field of trajectory planning, such as predicting the probability distribution of trajectories in two-dimensional environments through convolutional neural networks (CNN). However, the environmental complexity of three-dimensional space has increased significantly, and traditional methods face the following challenges: first, the computational efficiency is low, the sampling density of three-dimensional space increases exponentially, and uniform sampling leads to resource waste; second, the adaptability to dynamic obstacles is poor, and traditional algorithms are difficult to update trajectories in real time to cope with moving obstacles; third, the trajectory quality is unstable, the initial trajectory is tortuous, and the optimization process is time-consuming. Therefore, an efficient and adaptive three-dimensional trajectory planning method is urgently needed. Summary of the invention

[0003] In order to solve the above technical problems existing in the prior art, the present invention proposes a robot three-dimensional trajectory planning method based on deep neural network, which is applicable to the fields of industrial robots, service robots and autonomous driving, and realizes efficient and safe trajectory planning in complex environments. The specific technical solution is as follows:

[0004] A robot three-dimensional trajectory planning method based on a deep neural network, comprising:

[0005] Step S1: construct an environmental map and encode environmental information in the form of a digital grid;

[0006] Step S2: predicting the trajectory probability distribution in the environment map through a deep neural network model;

[0007] Step S3: generating non-uniform sampling points based on the trajectory probability distribution, while retaining some global uniform sampling points to form a hybrid sampling strategy;

[0008] Step S4: using an improved RRT* algorithm to perform trajectory search, wherein the improved RRT* algorithm optimizes the trajectory cost by dynamically adjusting the step size and rewiring mechanism;

[0009] Step S5: Update the environment information in real time and re-predict the trajectory probability distribution, triggering local trajectory replanning to deal with dynamic obstacles.

[0010] Furthermore, the input of the deep neural network model includes the environment map, the robot safety distance C and the step parameter S, and the output is the probability value of each grid unit being on the optimal trajectory.

[0011] Furthermore, the deep neural network model includes:

[0012] Encoder module: ResNet50 is used to extract multi-scale environmental features, and the feature pyramid network FPN is combined to enhance perception. ;

[0013] Attribute fusion module: The robot safety distance C and step length parameter S are mapped into feature vectors through the fully connected layer and concatenated with the map features. The expression is:

[0014] ,

[0015] Among them, MLP represents the fully connected layer mapping robot parameter feature vector;

[0016] Decoder module: progressively upsamples via transposed convolutions, outputting a probability map with the same resolution as the input map.

[0017] Furthermore, the training process of the deep neural network model includes:

[0018] Use the RRT* algorithm to generate optimal trajectories in a simulation environment as training data;

[0019] Each track is expanded to generate a label map with a width of 3 pixels, where the probability of the center area of ​​the track is 1 and the probability of the extended areas on both sides decreases;

[0020] The weighted cross entropy loss function is used to enhance the prediction accuracy of key areas of the trajectory;

[0021] The weighted cross entropy loss function is:

[0022] ,

[0023] in, Represents the label map, generated by RRT* trajectory expansion; represents the weight matrix, the trajectory center , decreasing on both sides to .

[0024] Furthermore, the hybrid sampling strategy in step S3 is specifically: selecting sampling points from hotspots with a probability value greater than 0.5 with a probability of 50%, and uniformly sampling from the entire environment space with a probability of 50%, that is:

[0025] .

[0026] Furthermore, the dynamic step size adjustment mechanism in step S4 includes: using a minimum step size of 1 in the hotspot area to refine the search, and using a maximum step size of 4 in the low probability area to accelerate the exploration, that is:

[0027] ,in, is the step length.

[0028] Furthermore, the local trajectory replanning in step S5 includes: when an environmental change is detected, only the tree nodes in the affected area are pruned, and the pruning condition is that if an obstacle change is detected in the area Ω, the tree node set is deleted ; Based on the updated probability distribution, new trajectory nodes are generated in the hotspot area first, that is, new nodes Priority from sampling.

[0029] Furthermore, the deep neural network model is a 3D convolutional network, and the expression is:

[0030] ,

[0031] in, is the voxelized environment representation with a resolution of 3×3×3; C is the robot safety distance, and S is the step size parameter.

[0032] Beneficial effects of the present invention:

[0033] This invention applies deep neural networks to probability modeling in three-dimensional trajectory planning for the first time, solving the problem of inefficient sampling in high-dimensional space in traditional methods; combining non-uniform sampling with the asymptotic optimality of RRT*, balancing exploration and utilization, reducing 90% of redundant nodes; integrating robot safety distance, step length and other parameters to generate safe trajectories that meet dynamic constraints, achieving multi-objective optimization and avoiding local optimality. This method significantly improves planning efficiency in complex environments, shortens the average trajectory length, reduces computing resource consumption, and achieves efficient trajectory planning in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a flowchart of a robot three-dimensional trajectory planning method based on a deep neural network according to an embodiment of the present invention;

[0035] Figure 2 is a schematic diagram of the structure of the deep neural network in step S2 of this embodiment;

[0036] Figure 3 Schematic diagram of trajectory selection between sampling points in steps S4-S5 of this embodiment. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0038] like Figure 1 As shown, this embodiment discloses a robot three-dimensional trajectory planning method based on a deep neural network, which specifically includes the following steps:

[0039] Step S1: Construct an environment map and encode the environment information into a two-dimensional or three-dimensional digital grid, where the value of each grid cell represents free space, obstacle, start point or end point, as shown below:

[0040] ,

[0041] in, : free space; :obstacle; :starting point; :end.

[0042] Step S2: Predict the trajectory probability distribution through a deep neural network model. The input of the deep neural network model includes the environment map, the robot safety distance C and the step parameter S. The output is the probability value of each grid cell being on the optimal trajectory, which is expressed as follows:

[0043] ,

[0044] in, : Sigmoid function, output probability value ; : Deep neural network model, parameter is θ; C: robot safety distance; S: step size parameter.

[0045] The training process of the deep neural network model includes:

[0046] Use the RRT* algorithm to generate a large number of optimal trajectories in a simulation environment as training data;

[0047] Each track is expanded to generate a label map with a width of 3 pixels, where the probability of the center area of ​​the track is 1 and the probability of the extended areas on both sides decreases;

[0048] The weighted cross entropy loss function is used to enhance the prediction accuracy of key areas of the trajectory, where the loss function is:

[0049] ,

[0050] in, Represents the label map, generated by RRT* trajectory expansion; represents the weight matrix, the trajectory center , decreasing on both sides to .

[0051] like Figure 2 As shown, the structure of the deep neural network model includes:

[0052] Encoder module: ResNet50 is used to extract multi-scale environmental features , combined with the feature pyramid network FPN to enhance perception, we get ; Attribute fusion module: Map the robot parameters C and S into feature vectors through the fully connected layer and concatenate them with the map features. The expression is:

[0053] ,

[0054] Among them, MLP represents the fully connected layer mapping robot parameter feature vector;

[0055] Decoder module: progressively upsamples via transposed convolutions and outputs a probability map with the same resolution as the input map.

[0056] Step S3: generating non-uniform sampling points based on the probability distribution, while retaining some global uniform sampling points to form a mixed sampling strategy.

[0057] The hybrid sampling strategy is specifically as follows: select sampling points from hotspots with a probability value greater than 0.5 with a probability of 50%; avoid ineffective exploration of low-probability areas by rejecting sampling; and uniformly sample from the entire environment space with a probability of 50%, that is:

[0058] .

[0059] Step S4: Use the improved RRT* algorithm to search for trajectories and optimize the trajectory cost by dynamically adjusting the step size mechanism and the rewiring mechanism.

[0060] The dynamic step size adjustment mechanism includes: using a minimum step size of 1 in the hotspot area to refine the search; using a maximum step size of 4 in the low probability area to accelerate the exploration, that is:

[0061] ,in, is the step length.

[0062] Step S5: Update the environment information in real time and re-predict the probability distribution, triggering local trajectory replanning to deal with dynamic obstacles, such as Figure 3 shown.

[0063] The local trajectory replanning includes: when an environmental change is detected, only the tree nodes in the affected area are pruned, and the condition is that if an obstacle change is detected in the area Ω, the tree node set is deleted Based on the updated probability distribution, new trajectory nodes are generated in the hotspot area first, that is, new nodes Priority from sampling.

[0064] The method of the present invention is also applicable to a three-dimensional environment, where the environmental information is represented by a point cloud or voxel grid map, and the deep neural network model can be replaced by a 3D convolutional network, expressed as:

[0065] ,

[0066] in, It is a voxelized environment representation with a resolution of 3×3×3.

[0067] Based on the above method, this embodiment also discloses a robot three-dimensional trajectory planning system, including: an input module, receiving three-dimensional environmental information such as point cloud / voxel map, starting point and end point coordinates, robot dynamic parameters such as maximum speed, safety radius, etc.; a neural network module: using 3D CNN or point cloud network to extract environmental features and predict the probability distribution of trajectory key points in three-dimensional space; a sampling optimization module: generating non-uniform sampling points based on probability distribution, and combining the improved RRT* algorithm for trajectory search and optimization.

[0068] In summary, the present invention is a robot three-dimensional trajectory planning method based on deep neural network, which predicts the probability distribution of trajectories in the environment by training the neural network model, guides the non-uniform sampling process, and combines the improved RRT* algorithm to achieve efficient trajectory search and dynamic optimization. This method can significantly improve planning efficiency in complex environments, shorten the average trajectory length, and reduce computing resource consumption.

[0069] The above is only a preferred implementation case of the present invention and does not limit the present invention in any form. Although the implementation process of the present invention is described in detail above, for those familiar with the art, they can still modify the technical solutions recorded in the above examples, or replace some of the technical features therein with equivalents. All modifications, equivalent replacements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A robot three-dimensional trajectory planning method based on deep neural network, characterized in that: include: Step S1: construct an environmental map and encode environmental information in the form of a digital grid; Step S2: predicting the trajectory probability distribution in the environment map through a deep neural network model; Step S3: generating non-uniform sampling points based on the trajectory probability distribution, while retaining some global uniform sampling points to form a hybrid sampling strategy; Step S4: using an improved RRT* algorithm to perform trajectory search, wherein the improved RRT* algorithm optimizes the trajectory cost by dynamically adjusting the step size and rewiring mechanism; Step S5: Update the environment information in real time and re-predict the trajectory probability distribution, triggering local trajectory replanning to deal with dynamic obstacles.

2. The robot three-dimensional trajectory planning method according to claim 1, characterized in that: The input of the deep neural network model includes the environment map, the robot safety distance C and the step size parameter S, and the output is the probability value of each grid unit being on the optimal trajectory.

3. The robot three-dimensional trajectory planning method according to claim 2, characterized in that: The deep neural network model includes: Encoder module: ResNet50 is used to extract multi-scale environmental features, and the feature pyramid network FPN is combined to enhance perception. ; Attribute fusion module: The robot safety distance C and step length parameter S are mapped into feature vectors through the fully connected layer and concatenated with the map features. The expression is: ， Among them, MLP represents the fully connected layer mapping robot parameter feature vector; Decoder module: progressively upsamples via transposed convolutions, outputting a probability map with the same resolution as the input map.

4. The robot three-dimensional trajectory planning method according to claim 2, characterized in that: The training process of the deep neural network model includes: Use the RRT* algorithm to generate optimal trajectories in a simulation environment as training data; Each track is expanded to generate a label map with a width of 3 pixels, where the probability of the center area of ​​the track is 1 and the probability of the extended areas on both sides decreases; The weighted cross entropy loss function is used to enhance the prediction accuracy of key areas of the trajectory; The weighted cross entropy loss function is: ， in, Represents the label map, generated by RRT* trajectory expansion; represents the weight matrix, the trajectory center , decreasing on both sides to .

5. The robot three-dimensional trajectory planning method according to claim 1, characterized in that: The hybrid sampling strategy in step S3 is specifically: select sampling points from hotspots with a probability value greater than 0.5 with a probability of 50%, and perform uniform sampling from the entire environment space with a probability of 50%, that is: 。 6. The robot three-dimensional trajectory planning method according to claim 1, characterized in that: The dynamic step size adjustment mechanism in step S4 includes: using a minimum step size of 1 in the hotspot area to refine the search, and using a maximum step size of 4 in the low probability area to accelerate the exploration, that is: ,in, is the step length.

7. The robot three-dimensional trajectory planning method according to claim 6, characterized in that: The local trajectory replanning in step S5 includes: when an environmental change is detected, only the tree nodes in the affected area are pruned, and the pruning condition is that if an obstacle change is detected in the area Ω, the tree node set is deleted ; Based on the updated probability distribution, new trajectory nodes are generated in the hotspot area first, that is, new nodes Priority from sampling.

8. The robot three-dimensional trajectory planning method according to claim 1, characterized in that: The deep neural network model is a 3D convolutional network, expressed as: ， in, is the voxelized environment representation with a resolution of 3×3×3; C is the robot safety distance, and S is the step size parameter.

Images