A method and system for robot arm motion planning based on deep neural networks

Abstract

The application provides a kind of mechanical arm motion planning method and system based on deep neural network.The method includes processing obstacle space point cloud data collected by depth camera to obtain processed point cloud data;Convert the processed point cloud data into octomap chart and import into moveit mechanical arm workspace;In moveit mechanical arm workspace, the complete collision-free path from the initial configuration state to the target configuration state is output by the mechanical arm motion planning algorithm based on deep neural network;Optimize the complete collision-free path to generate the optimal complete collision-free path;Control the mechanical arm to move according to the generated optimal complete collision-free path.The application proposes that the deep neural network predicts the configuration state of the mechanical arm at a certain time, reduces the calculation amount in the process of mechanical arm obstacle avoidance planning through collision detection and bidirectional iterative exploration, greatly speeds up the path planning speed, shortens the path planning time, and at the same time ensures that the planned path is close to the optimal path.

Description

Technical Field

[0001] This invention relates to the field of computer technology, specifically to a method and system for planning the motion of a robotic arm based on deep neural networks. Background Technology

[0002] Path planning for robotic arms in space has always been a challenging problem due to their complex configuration. With the development of robotics technology, robotic arm applications require fast and efficient motion planning algorithms. However, as the dimensionality of the robotic arm's workspace increases, the computational cost of collision detection routines also increases, making existing motion planning methods less effective.

[0003] For example, when it comes to motion planning in the space of a multi-degree-of-freedom robotic arm, traditional sampling-based algorithms will lead to a significant increase in the computational cost of collision detection routines as the dimension of the robotic arm's workspace increases. This makes the motion planning method less effective, the convergence to the optimal solution is slow, and the quality of the initial path cannot be guaranteed.

[0004] For example, adaptive sampling is one method to accelerate sampling-based motion planning algorithms by sampling specific regions of the configuration space that are more likely to contain optimal path solutions. Although there are many types of adaptive sampling algorithms, they all rely on hand-designed heuristics. Furthermore, their performance degrades significantly in high-dimensional spaces.

[0005] In summary, current neural network-based robotic arm motion planning algorithms either predict the free configuration space of the robotic arm or predict the configuration state of the robotic arm through neural networks. Although these existing methods reduce path planning time to some extent, they still require a certain amount of sampling and computation. Moreover, none of these methods include path optimization, and the quality of the planned path is difficult to guarantee. Summary of the Invention

[0006] In view of this, the present invention provides a method and system for robotic arm motion planning based on deep neural networks, so as to at least solve the above-mentioned problems.

[0007] This invention provides a robotic arm motion planning method based on deep neural networks, comprising: processing obstacle space point cloud data acquired by a depth camera to obtain processed point cloud data; converting the processed point cloud data into an octomap and importing it into the MoveIt robotic arm workspace; outputting a collision-free complete path from the initial configuration state to the target configuration state in the MoveIt robotic arm workspace using a robotic arm motion planning algorithm based on deep neural networks; optimizing the collision-free complete path to generate an optimal collision-free complete path; and controlling the robotic arm to move according to the generated optimal collision-free complete path.

[0008] In one implementation, the MoveIt robotic arm workspace outputs a collision-free complete path from the initial configuration state to the target configuration state using a deep neural network-based robotic arm motion planning algorithm. This includes: Step a, given a set of initial and target configuration states for the robotic arm, predicting the next configuration state of the robotic arm using a trained deep neural network model, and recording this next configuration state as an intermediate configuration state; Step b, checking for collisions between the initial and intermediate configuration states using a collision checker. If there are no collisions between the initial and intermediate configuration states, the intermediate configuration state is written into the path. If there are collisions between the initial and intermediate configuration states, the process returns and re-plans the path; Step c, after detecting no collisions between the initial and intermediate configuration states, checking for collisions between the intermediate and target configuration states. If there are no collisions between the intermediate and target configuration states, the path exploration ends, and a collision-free complete path is output. If there are collisions between the intermediate and target configuration states, the process iterates, recording the intermediate configuration state as the initial configuration state, and repeating steps a and c to continue path planning until a collision-free complete path from the initial configuration state to the target configuration state is found.

[0009] In another implementation, the input to the deep neural network model is a set of initial and target configuration states of the robotic arm, and the output of the deep neural network model is the configuration state of the robotic arm at the next moment.

[0010] In another implementation, the loss function of the deep neural network model is:

[0011]

[0012] Where, N p It is the number of path samples in the dataset. It is a collection of path samples.

[0013] In another implementation, optimizing the collision-free complete path to generate the optimal collision-free complete path includes: re-sampling the collision-free complete path at equal intervals within the joint space of the robotic arm, dividing the entire collision-free complete path into several segments; connecting all the sampling points obtained from the equal-interval sampling in pairs, performing collision detection between every two sampling point connections, omitting the path nodes between the detected collision-free sampling points, connecting the remaining sampling points, and obtaining the optimal collision-free complete path.

[0014] According to a second aspect of the present invention, a robotic arm motion planning system based on a deep neural network is provided, comprising: a robotic arm, a data processing module, a conversion module, a path planning module, an optimization module, and a movement module, wherein each module is communicatively connected to the robotic arm; the data processing module is used to process obstacle spatial point cloud data acquired by a depth camera to obtain processed point cloud data; the conversion module is used to convert the processed point cloud data into an octomap and import it into the MoveIt robotic arm workspace; the path planning module is used to output a collision-free complete path from the initial configuration state to the target configuration state in the MoveIt robotic arm workspace using a robotic arm motion planning algorithm based on a deep neural network; the optimization module is used to optimize the collision-free complete path to generate an optimal collision-free complete path; and the movement module is used to control the robotic arm to move according to the generated optimal collision-free complete path.

[0015] In summary, this invention proposes a deep neural network to predict the configuration state of the robotic arm at a certain moment. Through collision detection and bidirectional iterative exploration, it further reduces the amount of computation in the obstacle avoidance planning process of the robotic arm, greatly accelerates the path planning speed, shortens the path planning time, and ensures that the planned path is close to the optimal path. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of the robotic arm motion planning method based on deep neural networks of the present invention.

[0017] Figure 2 To and Figure 1 The working principle diagram of the corresponding robotic arm motion planning method based on deep neural networks;

[0018] Figure 3 The flowchart shows the motion planning algorithm for a robotic arm based on a deep neural network.

[0019] Figure 4 This is a schematic diagram of a deep neural network model structure;

[0020] Figure 5 This is a schematic diagram of the path optimization part in the motion planning algorithm for a robotic arm based on a deep neural network.

[0021] Figure 6 This is a schematic diagram of the bidirectional exploration process in the motion planning algorithm for a robotic arm based on a deep neural network. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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 are within the scope of protection of the present invention.

[0023] See Figure 1 This invention provides a robotic arm motion planning method based on deep neural networks, comprising:

[0024] Step S1: Process the obstacle spatial point cloud data acquired by the depth camera to obtain processed point cloud data;

[0025] Step S2: Convert the processed point cloud data into an octomap and import it into the Moveit robotic arm workspace;

[0026] Step S3: In the MoveIt robotic arm workspace, output a collision-free complete path from the initial configuration state to the target configuration state using a robotic arm motion planning algorithm based on a deep neural network;

[0027] Step S4: Optimize the collision-free complete path to generate the optimal collision-free complete path;

[0028] Step S5: Control the robotic arm to move along the generated optimal collision-free complete path.

[0029] Further, see Figure 2 First, the obstacle space point cloud data is acquired using a depth camera, processed, and then converted into an octomap image for import into the robotic arm's workspace. A deep neural network model (FPPNet) is trained, and a motion planner based on this deep neural network (with the motion planning algorithm integrated into the planner) ultimately generates a complete path (optimal collision-free complete path) from the starting point to the target state. Finally, the robotic arm is controlled to move along this generated optimal collision-free complete path.

[0030] For example, in a ROS environment, a robotic arm is controlled via MoveIt. A depth camera collects and processes obstacle space data, which is then converted into an octomap and imported into the MoveIt robotic arm's workspace to update the obstacles within the workspace. A motion planning algorithm then effectively enables the vision-based robotic arm to autonomously avoid obstacles and plan its path. This process mainly includes the following three steps:

[0031] (1) In the ROS environment, the robotic arm is controlled by moveit;

[0032] (2) Construct an octomap for collision detection and obstacle avoidance;

[0033] (3) Use a depth camera to perceive environmental data and import the octomap into the UR5 workspace through scene planning_scene.

[0034] In addition, the calibration of the robotic arm system adopts an eye-out calibration scheme. By printing a calibration board, the hand-eye calibration of the UR5 robotic arm and the RealSense depth camera is completed by calling the aruco library in the ROS environment.

[0035] In one implementation, the MoveIt robotic arm workspace outputs a collision-free complete path from the initial configuration state to the target configuration state using a deep neural network-based robotic arm motion planning algorithm. This includes: Step a, given a set of initial and target configuration states for the robotic arm, predicting the next configuration state of the robotic arm using a trained deep neural network model, and recording this next configuration state as an intermediate configuration state; Step b, checking for collisions between the initial and intermediate configuration states using a collision checker. If there are no collisions between the initial and intermediate configuration states, the intermediate configuration state is written into the path. If there are collisions between the initial and intermediate configuration states, the process returns and re-plans the path; Step c, after detecting no collisions between the initial and intermediate configuration states, checking for collisions between the intermediate and target configuration states. If there are no collisions between the intermediate and target configuration states, the path exploration ends, and a collision-free complete path is output. If there are collisions between the intermediate and target configuration states, the process iterates, recording the intermediate configuration state as the initial configuration state, and repeating steps a and c to continue path planning until a collision-free complete path from the initial configuration state to the target configuration state is found.

[0036] In another implementation, the input to the deep neural network model is a set of initial and target configuration states of the robotic arm, and the output of the deep neural network model is the configuration state of the robotic arm at the next moment.

[0037] In another implementation, the loss function of the deep neural network model is:

[0038]

[0039] Where, N p It is the number of path samples in the dataset. It is a collection of path samples.

[0040] In another implementation, optimizing the collision-free complete path to generate the optimal collision-free complete path includes: re-sampling the collision-free complete path at equal intervals within the joint space of the robotic arm, dividing the entire collision-free complete path into several segments; connecting all the sampling points obtained from the equal-interval sampling in pairs, performing collision detection between every two sampling point connections, omitting the path nodes between the detected collision-free sampling points, connecting the remaining sampling points, and obtaining the optimal collision-free complete path.

[0041] For example, the principle of a robotic arm motion planning algorithm based on deep neural networks is as follows: Figure 3 As shown, given a set of initial and target configuration states for a robotic arm, a trained deep neural network model (FPPNet) predicts the next configuration state of the robotic arm, which is denoted as the intermediate configuration state. A collision checker then checks for collisions between the initial configuration state and the intermediate state. If no collision occurs, the intermediate configuration state is written into the path; if a collision occurs, the process returns, and path planning is restarted. Next, the process checks for collisions between the intermediate state and the target configuration state. If no collision occurs, path exploration ends, and the path is output; if a collision occurs, the process iterates, recording the intermediate state as the initial configuration state, returning to the beginning, and continuing path planning until a collision-free complete path from the initial configuration state to the target state is found. Finally, the output collision-free complete path undergoes path optimization to reduce path cost and smooth the trajectory, ultimately yielding a near-optimal complete collision-free path.

[0042] Figure 4 This is a schematic diagram of the deep neural network model (FPPNet). The FPPNet model is the core of the entire algorithm. The model input consists of the initial and target configuration states of the robotic arm, with an input dimension of 12 (the algorithm experiments all use a 6-DOF UR5 robotic arm). The model predicts the next configuration state of the robotic arm. When no collision occurs in the robotic arm's workspace, the current output configuration is written into the path and marked as the initial configuration. This initial configuration is then combined with the target configuration state and input back into the FPPNet model for prediction, iteratively deriving the complete path.

[0043] FPPNet network input: A set of initial and target configuration states for the robotic arm;

[0044] FPPNet network output: The configuration status of the robotic arm at the next moment.

[0045] FPPNet network structure: It adopts a 14-layer DNN network, where each hidden layer consists of a linear layer, PReLU, and Dropout, except for the last hidden layer which does not include Dropout. The last layer is the output layer, which takes the hidden units from the previous layer and transforms them into the required output dimension, which is consistent with the configuration dimension of the robotic arm.

[0046] Loss function: The training objective of the neural network is to minimize the robot configuration state predicted by the network. And the actual robot configuration state σ in the dataset i+1 The mean squared error (MSE) between them. The form of the neural network loss function is:

[0047]

[0048] Where, N p It is the number of path samples in the dataset. It is a collection of path samples.

[0049] Dataset Acquisition: To train the network, approximately 50,000 path samples were generated in the workspace using the RRT* planner for different planning tasks (for the UR5 robotic arm, the default RRT* and Moveit parameter settings in OMPL were used to generate the paths). The data was preprocessed by decomposing a successfully planned path into several path nodes, and then combining the configuration state of the robotic arm at each path node with the target configuration state of that path as the input to the neural network model.

[0050] Model training: The training objective is to minimize the robot configuration state predicted by the network. And the actual robot configuration state σ in the dataset i+1 The mean squared error between the values ​​is between 1 and 10. The training batch size of the FPPNet network is 32, and the learning rate is 1×10. -3 Up to 1×10 -6 Fine-tuning is required between different scenarios, and approximately 200 epochs are needed to complete the training.

[0051] It should be understood that the collision detection in the entire robotic arm motion planning algorithm is divided into two parts:

[0052] Firstly, after the model outputs the robotic arm configuration state, it checks whether the output state collides with obstacles in the workspace, and whether there is a collision on the path from the initial configuration to the output configuration. Once no collision is detected, the path can be written.

[0053] Secondly, the algorithm also includes collision detection between the current configuration state output by the robotic arm and the target configuration state. If there is no collision, there is no need to iterate further, and the complete path can be output directly.

[0054] Given that the model's predictions can be significantly affected by training accuracy, simply relying on model inference to continuously approach the target configuration will not be effective. Implementing dual collision detection can effectively reduce the number of model inferences and lower path planning time. Issues such as unstable output path cost and insufficient trajectory smoothness can be addressed in the subsequent path optimization section.

[0055] like Figure 5 The diagram shows the path optimization part. The path optimization part of the robotic arm motion planning algorithm processes the collision-free complete path from the initial to the target state obtained from the previous model iterations. Because the model prediction has a certain degree of instability, the resulting path cost is also unstable, requiring cost optimization and trajectory smoothing operations on the output path. The specific optimization process is as follows:

[0056] Within the joint space of the robotic arm, the collision-free complete path output by the previous step of the robotic arm motion planning algorithm is resampled at equal intervals, dividing the entire collision-free complete path into several segments. Then, all sampled points are connected pairwise, and collision detection is performed between each pair of sampled points. When no collision is detected between two sampled points, the path nodes between these two points can be omitted. Finally, connecting the remaining sampled points yields the optimized path, which is the optimal collision-free complete path. Path optimization can yield a collision-free and near-optimal path.

[0057] like Figure 6 The diagram illustrates the bidirectional exploration process of the robotic arm motion planning algorithm. This algorithm involves simultaneous exploration from the initial configuration state towards the target configuration state and from the target configuration state towards the initial configuration state. The path exploration proceeds concurrently in both directions, and the configuration states output by the model are continuously added to the path sets in both directions. Thus, when the algorithm checks for collisions between intermediate configuration states and the target configuration state, it essentially checks for collisions between two intermediate configuration states in each of the two exploration directions. If no collision is detected between the two intermediate configuration states, the two output path sets are integrated according to the direction from the initial configuration state to the target configuration state, ultimately outputting a collision-free complete path. Compared to unidirectional path exploration, the bidirectional exploration method effectively reduces path planning time.

[0058] Experimental verification:

[0059] By setting different scenarios and tasks on simulation and actual platforms, the deep neural network-based robotic arm motion planning algorithm (FPPNet algorithm) mentioned in this invention has shown superior performance compared to existing algorithms (MPNet algorithm, RRT* algorithm).

[0060] Analysis of the experimental data shows that the FPPNet algorithm has a significant advantage in path planning time compared to the RRT* and MPNet algorithms, and the FPPNet algorithm can reduce the planning time by about 10 times compared to the MPNet algorithm. The path cost of the FPPNet algorithm is close to that of the RRT* algorithm and slightly lower than that of the MPNet algorithm.

[0061] In summary, this invention proposes a deep neural network to predict the configuration state of the robotic arm at a certain moment. Through collision detection and bidirectional iterative exploration, it further reduces the amount of computation in the obstacle avoidance planning process of the robotic arm, greatly accelerates the path planning speed, shortens the path planning time, and ensures that the planned path is close to the optimal path.

[0062] This invention also provides a robotic arm motion planning system based on a deep neural network, comprising:

[0063] The system includes a robotic arm, a data processing module, a conversion module, a path planning module, an optimization module, and a movement module, with each module communicating with the robotic arm.

[0064] The data processing module is used to process the obstacle spatial point cloud data acquired by the depth camera to obtain the processed point cloud data.

[0065] The conversion module is used to convert the processed point cloud data into an octomap image and import it into the Moveit robotic arm workspace.

[0066] The path planning module is used to output a collision-free complete path from the initial configuration state to the target configuration state in the moveit robot arm workspace using a robot arm motion planning algorithm based on a deep neural network.

[0067] The optimization module is used to optimize the collision-free complete path and generate the optimal collision-free complete path;

[0068] The movement module controls the robotic arm to move along the generated optimal collision-free complete path.

[0069] In summary, this invention proposes a deep neural network to predict the configuration state of the robotic arm at a certain moment. Through collision detection and bidirectional iterative exploration, it further reduces the amount of computation in the obstacle avoidance planning process of the robotic arm, greatly accelerates the path planning speed, shortens the path planning time, and ensures that the planned path is close to the optimal path.

[0070] The robotic arm motion planning system based on deep neural networks in this embodiment is used to implement the corresponding methods in the aforementioned multiple method embodiments and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0071] It should be noted that the present invention can be a method, system, apparatus, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the present invention.

[0072] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0073] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0074] The computer program instructions used to perform the operations of this invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, Python, etc., and conventional procedural programming languages ​​such as "C" or similar languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing state information from the computer-readable program instructions. This electronic circuitry can execute the computer-readable program instructions to implement various aspects of the invention.

[0075] Various aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0076] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0077] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0078] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. It will be known to those skilled in the art that implementation in hardware, implementation in software, and implementation using a combination of software and hardware are equivalent.

[0079] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, and are not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein. The scope of the invention is defined by the appended claims.

Claims

1. A robotic arm motion planning method based on deep neural networks, characterized in that, include: The spatial point cloud data of obstacles acquired by the depth camera is processed to obtain the processed point cloud data; The processed point cloud data is converted into an octomap and imported into the Moveit robotic arm workspace. In the MoveIt robotic arm workspace, a collision-free complete path from the initial configuration state to the target configuration state is output using a deep neural network-based robotic arm motion planning algorithm. This includes: Step a, given a set of initial and target configuration states for the robotic arm, the next configuration state of the robotic arm is predicted using a trained deep neural network model, and this next configuration state is recorded as an intermediate configuration state; Step b, a collision checker checks for collisions between the initial and intermediate configuration states. If there are no collisions between the initial and intermediate configuration states, the intermediate configuration state is written into the path. If there are collisions between the initial and intermediate configuration states, the process returns and path planning is restarted; Step c, when no collisions are found between the initial and intermediate configuration states, a collision check is performed between the intermediate and target configuration states. If there are no collisions between the intermediate and target configuration states, path exploration ends, and a collision-free complete path is output. If there are collisions between the intermediate and target configuration states, the process iterates, recording the intermediate configuration state as the initial configuration state, and repeating steps a and c to continue path planning until a collision-free complete path from the initial configuration state to the target configuration state is found. Optimize the collision-free complete path to generate the optimal collision-free complete path; Control the robotic arm to move along the generated optimal collision-free complete path.

2. The method according to claim 1, characterized in that, The input to the deep neural network model is a set of initial and target configuration states of the robotic arm, and the output of the deep neural network model is the configuration state of the robotic arm at the next moment.

3. The method according to claim 2, characterized in that, The loss function for a deep neural network model is: in, It is the number of path samples in the dataset. It is a collection of path samples. Configure the robot's state according to the network's predictions. Configure the actual robot status in the dataset.

4. The method according to claim 1, characterized in that, The optimization of the collision-free complete path to generate the optimal collision-free complete path includes: Within the joint space of the robotic arm, the collision-free complete path is resampled at equal intervals, and the entire collision-free complete path is divided into several segments; Connect all the sampling points obtained by equal-interval sampling in pairs, perform collision detection between every two sampling points, omit the path nodes between the sampling points that are detected without collision, connect the remaining sampling points, and obtain the optimal collision-free complete path.

5. A robotic arm motion planning system based on deep neural networks, characterized in that, include: The system includes a robotic arm, a data processing module, a conversion module, a path planning module, an optimization module, and a movement module, with each module communicating with the robotic arm. The data processing module is used to process the obstacle spatial point cloud data acquired by the depth camera to obtain the processed point cloud data. The conversion module is used to convert the processed point cloud data into an octomap image and import it into the Moveit robotic arm workspace. The path planning module is used to output a collision-free complete path from the initial configuration state to the target configuration state in the Moveit robotic arm workspace using a deep neural network-based robotic arm motion planning algorithm. The steps include: Step a: Given a set of initial and target configuration states for the robotic arm, predict the next configuration state of the robotic arm using a trained deep neural network model, and record this next configuration state as an intermediate configuration state; Step b: Check for collisions between the initial and intermediate configuration states using a collision checker. If there are no collisions between the initial and intermediate configuration states, write the intermediate configuration state into the path. If there are collisions between the initial and intermediate configuration states, return and re-perform path planning; Step c: When no collisions are found between the initial and intermediate configuration states, check for collisions between the intermediate and target configuration states. If there are no collisions between the intermediate and target configuration states, the path exploration ends, and a collision-free complete path is output. If there are collisions between the intermediate and target configuration states, iterate, record the intermediate configuration state as the initial configuration state, and repeat steps a and c to continue path planning until a collision-free complete path from the initial configuration state to the target configuration state is found. The optimization module is used to optimize the collision-free complete path and generate the optimal collision-free complete path; The movement module controls the robotic arm to move along the generated optimal collision-free complete path.

Images