A robotic arm autonomous assembly method and system suitable for unstructured environments

Abstract

The present invention discloses a robot arm autonomous assembly method suitable for unstructured environment, comprising: setting a total number of training rounds; the end of the robot arm moves with pre-planned path points, and the current state space information of each step is obtained as the input of the autonomous assembly network for training to form an initial autonomous assembly strategy model; the end of the robot arm obtains the optimal action in the current state through feed-forward calculation of the initial autonomous assembly strategy model, and the current state space information of each step is obtained as the input of the autonomous assembly network for updating the autonomous assembly strategy model, and after each set number of steps, Gaussian noise with different means is added to the probability of the autonomous assembly strategy model selecting an action; the training is stopped when the total number of training rounds is reached; the Gaussian noise of action selection is cancelled in the feed-forward calculation of the training model, the optimal action is selected according to the current state space information, and the action selection is stopped when the task completion identifier value changes, so as to complete the assembly task.

Description

Technical Field

[0001] The present invention belongs to the field of machine learning, and in particular relates to a method and system for autonomous assembly of a robotic arm suitable for unstructured environments. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

[0003] On industrial production lines, robotic arms are high-freedom actuators that can perform a variety of operations, such as plugging and unplugging, assembly, transportation, installation, and maintenance. At present, in high-precision assembly task scenarios, industrial robotic arms rely on the planning of fixed path points to complete assembly tasks. On the one hand, due to the fact that the robotic arm captures the object to be assembled and the repositioning of the robotic arm will produce certain errors in the actual production process, the robotic arm must have a certain self-adjustment ability. The existing robotic arm relies on fixed paths and target end positions and cannot reduce and avoid the impact and collision caused by errors through self-adjustment; on the other hand, environmental changes and task adjustments often occur in production and assembly tasks, and the fixed-path robotic arm control method cannot meet the dynamic response requirements of such changes. Therefore, the ever-expanding industrial production lines require robotic arm systems that can assemble autonomously.

[0004] In order to realize the autonomous assembly of the robot arm, the deep reinforcement learning technology of the online update model is used to train the model. Reinforcement learning abstracts the interaction process between the research target and the environment into a state space D(s,a,s',r,d), where s is the state quantity of interest in the environment where the research target is located at the current moment, a is the action taken by the research target at the current moment, s' is the state quantity of interest in the environment where the research target enters the next moment after taking action a, r is the reward score at the current moment, and d is the flag bit of whether the task is completed. The research target obtains information about the interaction with the environment through the state space and maximizes the cumulative reward value to complete the training of the policy model.

[0005] Since the robot assembly task is performed in a continuous state space, while the state space of traditional reinforcement learning is discrete. In order to solve this non-discrete state space problem, reinforcement learning introduces neural networks to deal with high-dimensional space problems. Designing a multi-layer neural network can not only ensure that the policy model makes correct decisions in a continuous state space, but also improve the decision-making accuracy of the intelligent agent in unknown states.

[0006] At present, it is difficult to apply deep reinforcement learning technology to real robotic arms. In different intelligent agents and different tasks, the input data dimensions and model hyperparameters of deep reinforcement learning have different adjustment standards. The normalization of input data and the magnitude of hyperparameters determine whether the policy model converges, which makes the robotic arm uncertain in the learning process. Summary of the invention

[0007] The technical problem to be solved by the present invention is: in order to overcome the deficiencies of the above-mentioned prior art, the present invention provides a method and system for autonomous assembly of a robot arm suitable for an unstructured environment, so that the robot arm can autonomously assemble an object to a specified position based on the end position and force sensor data;

[0008] The technical solution of the present invention is: a method for autonomous assembly of a robotic arm suitable for an unstructured environment, comprising the following steps:

[0009] S1. Set the total number of training rounds and allocate the number of rounds for the human-machine teaching phase and the self-learning phase;

[0010] S2: During the human-machine teaching stage, the end of the robot arm moves along the pre-planned path points, and the current state space information of each step is obtained as the input of the autonomous assembly network for training to form an initial autonomous assembly strategy model. After the human-machine teaching stage is completed, S3 is entered;

[0011] S3, enter the self-learning stage, the end of the robot arm obtains the optimal action in the current state through the feed-forward calculation of the initial autonomous assembly strategy model, obtains the current state space information of each step as the input of the autonomous assembly network, and is used to update the autonomous assembly strategy model. After every set number of steps, Gaussian noise with different means is added to the probability of the autonomous assembly strategy model selecting an action; when the total number of training rounds is reached, the training is stopped and enter S4;

[0012] S4. Cancel the Gaussian noise of action selection in the feedforward calculation of the training model, select the optimal action according to the current state space information, stop action selection when the task completion identifier value changes, and complete the assembly task.

[0013] Furthermore, in S1, the total number of training rounds is 6000, of which the number of random rounds is 1000, the number of learning rounds is 5000, and the number of training steps in each round is 100.

[0014] Furthermore, in S2, the state space information includes: three-dimensional coordinates of the end of the robot arm, three-dimensional position of the center point of the assembly object handle, three-dimensional force feedback value of the end of the robot arm, and coordinate value of the assembly end point.

[0015] Furthermore, in S2, the method for setting the pre-planned path points is as follows:

[0016] Set the step length of each step of the robot arm to 5mm, the starting position is the center point of the end of the robot arm gripper, and the end position is the center point of the bottom of the assembly slot. The starting and end positions will change randomly during each training;

[0017] According to the random assembly starting position and fixed assembly end position of each initialization scene, the deviation in the x and y directions is calculated, and the number of steps n required to align the center point is calculated by the deviation and step length. Then, in the first n steps, the center point of the assembly object handle (the assembly object refers to the object to be grasped by the robot arm, and the handle is set on the assembly object to facilitate the robot arm to grasp) is aligned with the assembly end position in the x and y directions;

[0018] Keep the x and y directions of the output action always at 0 and z at -1 until it is assembled in place.

[0019] Furthermore, in S2, the current state space information of each step is obtained as the input of the autonomous assembly network for training, including:

[0020] S21: Initialize the critic1, critic2, and actor network parameters, including the weights of each fully connected layer, the number of connected layers, and the bias value;

[0021] S22: Initialize three target networks and copy the network parameters of critic1, critic2, and actor to the target networks;

[0022] S23: Initialize the experience replay pool;

[0023] S24: Set the Gaussian noise distribution range for action selection and start the loop iteration of network update; select the optimal action according to the current network model and execute it, obtain the next state space information and the current reward score, and store the current state space information, the next state space information, the selected action, the reward score, and whether the task is completed as the experience of this step into the experience replay pool;

[0024] S25: Calculate the loss function of target network 1, target network 2, critic1, and critic2 through the Bellman equation, and update the network parameters of critic1 and critic2;

[0025] S26: The actor network uses gradient ascent to update network parameters after a delay;

[0026] S27: When the total number of training rounds reaches 6000, exit the loop and end the training.

[0027] Furthermore, in step S21, the actor network is the execution network, and the critic1 and critic2 networks are the evaluation networks. The execution network inputs the state quantity at the current moment and maps the action taken according to the state quantity; the evaluation network inputs the state quantity at the current moment and the action taken, maps the Q value (value) according to these quantities, and evaluates the performance of the execution network by the size of the Q value; the actor, critic1, and critic2 together constitute an autonomous assembly network.

[0028] Furthermore, in step S24, the method of selecting the optimal action according to the current network model is as follows:

[0029] The network model calculates the value at the current moment based on the input state space information, and calculates the action value with the highest probability at the current moment through a greedy algorithm.

[0030] Furthermore, in step S24, the calculation of the reward score is based on the reward function, and the reward function R is designed as follows:

[0031] R=r1+r2+r3

[0032] in, x1 and x2 are the coordinates of the center point of the assembly object handle and the coordinates of the assembly end point respectively; r2 = -max(f x , f y , f z ), f x 、f y 、f z are the three-dimensional force feedback values ​​at the end of the robotic arm; l1 and l2 are the coordinate values ​​of the center point of the assembly object handle and the coordinate value of the assembly starting point respectively; step represents the current step number of each round.

[0033] Furthermore, in S2, the movement of the end of the robot arm is achieved by solving the inverse kinematics of the robot arm; through a set of three-dimensional coordinate values ​​in a Cartesian coordinate system, the six-degree-of-freedom robot arm joint angles are solved in the order of joint 1, joint 3, joint 2, joint 4, joint 5, and joint 6; the movement mode of the end of the robot arm is to add the absolute coordinate value of the end position of the robot arm at the current moment plus the relative motion coordinate value predicted by the autonomous assembly network as the coordinate value of the end of the robot arm for the next step.

[0034] Furthermore, in S2, the range of motion of the end of the robot arm is limited:

[0035] The absolute coordinate value of the end position of the robot arm at the current moment is added to the relative motion coordinate value predicted by the autonomous assembly network. When the end position of the robot arm exceeds the threshold in the x, y, and z directions, the end of the robot arm maintains the position of the previous moment.

[0036] Furthermore, in S3, the method of adding Gaussian noise to the selected action is:

[0037] When the total number of training steps reaches 2000 rounds, Gaussian noise with a mean of 0 and a variance of 0.1 is added; when the total number of training steps reaches 3000 rounds, Gaussian noise with a mean of 0 and a variance of 0.3 is added; when the total number of training steps reaches 6000 rounds, Gaussian noise with a mean of 0 and a variance of 0.5 is added.

[0038] A robotic arm autonomous assembly system suitable for unstructured environments, including: a robotic arm, a three-dimensional force sensor and a control system;

[0039] The three-dimensional force sensor collects the forces in the x, y, and z directions applied to the end of the robot arm during the assembly process when the robot arm is grasping the assembly object, and inputs the force information into the autonomous assembly network on the control system; the autonomous assembly network is constructed using the above-mentioned robot arm autonomous assembly method suitable for unstructured environments;

[0040] The autonomous assembly network predicts the next position of the robot end according to the learned assembly strategy, and the control system sends the end position to the robot drive;

[0041] The robot arm performs actions according to the position sent by the control system.

[0042] The beneficial effects of the present invention compared with the prior art are:

[0043] (1) The present invention learns the robot arm assembly strategy through the autonomous assembly network, so that the robot arm can grab the object to be assembled and assemble it into the slot, while ensuring that the force applied during the assembly process remains within a safe threshold range when the external environment changes.

[0044] (2) The autonomous assembly network trained by the method of the present invention can adjust the motion path of the robot arm when the position of the assembly slot changes during the assembly process, so that the force on the object to be assembled is kept within a safe range and the object to be assembled and the assembly slot are not damaged. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0046] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0047] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0048] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0049] like Figure 1 As shown, this embodiment discloses a robotic arm autonomous assembly method suitable for unstructured environments, comprising:

[0050] (1) Set the total number of training rounds and allocate the number of rounds for the human-machine teaching phase and the self-learning phase;

[0051] (2) During the human-machine teaching stage, the end of the robot arm moves along the pre-planned path points, and obtains the current state space information of each step as the input of the autonomous assembly network for training to form an initial strategy model. After the human-machine teaching stage is completed, it enters step (3);

[0052] (3) After the initial strategy model is formed, the self-learning phase begins. The end of the robot arm obtains the optimal action in the current state through the feed-forward calculation of the initial strategy model, obtains the current state space information of each step as the input of the autonomous assembly network to update the model, and adds Gaussian noise with different means to the probability of the model selecting an action after a certain number of steps. When the total number of training rounds is reached, the training is stopped and the process goes to step (4);

[0053] (4) The Gaussian noise of action selection is eliminated in the feedforward calculation of the training model, the optimal action is selected according to the current state space information, and the action selection is stopped when the task completion identifier value changes, thus completing the assembly task.

[0054] The above-mentioned robot arm autonomous assembly method based on deep reinforcement learning is a model embedded method based on the TD3 algorithm, which enables the robot arm to autonomously select the next action of the robot arm based on the end position and force sensor. It can handle high-precision assembly tasks in an end-to-end manner.

[0055] The autonomous assembly network proposed in the above implementation example consists of two parts, namely, the execution network and the evaluation network. The execution network inputs the state quantity at the current moment and maps the actions taken according to the state quantity; the evaluation network inputs the state quantity at the current moment and the actions taken, maps the Q value (value) according to these quantities, and evaluates the performance of the execution network by the size of the Q value.

[0056] By learning assembly strategies in unstructured environments through an autonomous assembly network, the robotic arm can assemble objects into slots while ensuring that the forces remain within a safe threshold.

[0057] In order to accelerate the learning process, make the network pay more attention to task-related information and provide additional training signals, expert experience is introduced to assist training. The design method of expert experience is to design the shortest path point set for assembly based on the random initial end position and random slot position of the robot arm. At the same time, path constraints are added based on the feedback force information of the end force sensor. When the force in a certain direction is greater than the set safety threshold, the end of the robot arm is adjusted in the opposite direction of the force in that direction until the force is less than the safety threshold. Each time a round of assembly tasks is completed, the state space and action space information in the process is added to the experience pool as an expert experience for the robot arm to learn.

[0058] In the specific implementation process, a robotic arm system with a two-finger robotic gripper is used for operation. Since the state of the robotic arm conforms to the independent and identically distributed principle, the robotic arm assembly process can be expressed as a Markov decision process. Let t be the time step. In state s t Under the condition, the robot system collects the end position and force feedback information. Then, the robot arm follows the strategy π parameterized by θ θ (s t ) Select and implement the action a of the robot arm at this time step t (Specifically, it is to select the spatial position of the next step of the end of the robot arm). This strategy can be learned by training a neural network. After the robot arm selects an action through the strategy, it calculates the instantaneous reward r(s) for that time step. t , a t ), and update the state s of the next time step t+1 . Training ends to find an optimal strategy The optimal strategy is found by maximizing the cumulative reward value, which is the sum of the instantaneous rewards of the robot arm at all time steps in a training round.

[0059] The above framework provides a solution to this type of decision-making problem, but training is difficult due to the difficulty of data collection. Collecting a large amount of effective experience is crucial for the convergence of reinforcement learning models. In order to effectively train the network, the strategy of estimating the value function is estimated by minimizing the mean square error of the value of the execution network and the evaluation network.

[0060] Define the action representation of the assembly as a t , generating an action at each time step. The base of the robot is taken as the origin of the coordinate system, and the vertical upward direction of the base is the positive direction of the Z axis. For a specific assembly action, let the current end position of the robot be [x t ,y t , z t], the original action [Δx t , Δy t , Δz t ], then the next action to be performed by the robot arm is [x t +γΔx t ,y t +γΔy t , z t +γΔz t ], where γ is the distance coefficient of the actual environment.

[0061] The reward function of the robot arm autonomous assembly method is divided into three parts, which can ensure that the robot arm is assembled in place and the assembly force is protected in place. The reward function is defined as follows in the present invention:

[0062] R=r1+r2+r3

[0063] in, x1 and x2 are the coordinates of the center point of the assembly object handle and the coordinates of the assembly end point respectively; r2 = -max(f x , f y , f z ), f x 、f y 、f z are the three-dimensional force feedback values ​​at the end of the robotic arm; step is the number of steps executed in the current round, l1 and l2 are the coordinate values ​​of the center point of the assembly object handle and the coordinate value of the assembly starting point respectively.

[0064] In order to prevent the autonomous assembly network from falling into a local optimum during training, which causes the network to diverge and fail to converge. The range of motion of the end of the robotic arm is restricted, and the restriction ranges of 0.02m, 0.02m, and 0.27m are added to the x, y, and z directions of the center of the random assembly slot, respectively. When the next action selected by the robotic arm according to the strategy exceeds this range, the position of the previous step remains unchanged. Therefore, every time the robotic arm moves beyond the restricted range, the robotic arm does not respond to the action value, and due to the setting of r3 in the reward function, the instantaneous reward value of the robotic arm will always generate negative values, making the cumulative reward smaller and avoiding the divergence of the neural network training.

[0065] The deep Q network is used as an approximator of the Q function to train the autonomous assembly network. Specifically, the execution network is modeled as a three-layer fully connected layer network, where the input layer parameters are the state space dimensions (in this embodiment, s t The dimensions are set to 18) and 400, the hidden layer parameters are 400 and 300, the output layer parameters are 300 and the action space dimension (in this embodiment a tThe dimension is set to 3); the evaluation network is modeled into two 3-layer fully connected networks, where the input layer parameters are the state space dimension plus the action space dimension (set to 21 in this embodiment) and 400, the hidden layer parameters are 400 and 300, and the output layer parameters are 300 and 1. In this way, the execution network can be mapped from the state space to the action space, and the evaluation network can be mapped from the state space plus the action space to the Q value.

[0066] In order to accelerate the learning process, a distributed training framework is implemented in this embodiment. The self-assembly network model is distributed on different graphics cards. During the initialization process, each model loads the same model parameters. During the training process, the gradient of the loss function is calculated according to the state space information. After completing a round of training, each graphics card transmits its own gradient back together and takes the average value as the final gradient of this round of training. The model parameters are solved by the gradient and initialized as new model parameters at the beginning of the next round of training.

[0067] The system of this embodiment was evaluated in a simulation environment and a real environment. An assembly object of size 40.5 cm × 40.5 cm × 50 cm was placed into a slot of size 40 cm × 40 cm × 65 cm, and the forces in the three directions during the assembly process were ensured not to exceed 2 N. Two experiments were conducted:

[0068] Performance testing in a simulation environment to evaluate the contribution of each component of the system of the present invention to the overall performance;

[0069] It is demonstrated that the system of the present invention can be applied to real robotic autonomous assembly tasks.

[0070] The present invention uses a UR3 robot arm, a Robotiq six-dimensional force sensor and a Robotiq two-finger gripper to conduct robot testing in a real environment. During the test, the task success condition is set as the Euclidean distance between the center point of the object handle and the assembly end point is less than 1.5 cm. The method of the present invention has an autonomous assembly success rate of more than 99% in an unstructured environment, and the maximum assembly force is 1.75N.

[0071] In an embodiment, a robotic arm autonomous assembly system suitable for use in an unstructured environment includes: a robotic arm, a three-dimensional force sensor, and a control system;

[0072] The three-dimensional force sensor collects the forces in the x, y, and z directions applied to the end of the robot arm during the assembly process when the robot arm is grasping the assembly object, and inputs the force information into the autonomous assembly network; the autonomous assembly network is constructed using the above-mentioned robot arm autonomous assembly method suitable for unstructured environments;

[0073] The autonomous assembly network predicts the next position of the robot end according to the learned assembly strategy, and the control system sends the end position to the robot drive;

[0074] The robot arm selects and implements actions based on the prediction results.

[0075] When the system is working, specific steps of a robotic arm autonomous assembly method suitable for unstructured environments are described.

[0076] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0077] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

[0078] Parts of the present invention that are not described in detail belong to the well-known technologies of those skilled in the art.

Claims

1. A robotic arm autonomous assembly method suitable for unstructured environments, characterized in that: include: Set the total number of training rounds, and allocate the number of rounds for the human-machine teaching phase and the self-learning phase; In the human-machine teaching stage, the end of the robot arm moves along the pre-planned path points, and the current state space information of each step is obtained as the input of the autonomous assembly network for training to form an initial autonomous assembly strategy model; Entering the self-learning stage, the end of the robot arm obtains the optimal action in the current state through the feed-forward calculation of the initial autonomous assembly strategy model, obtains the current state space information of each step as the input of the autonomous assembly network, and is used to update the autonomous assembly strategy model. After every set number of steps, Gaussian noise with different means is added to the probability of the autonomous assembly strategy model selecting an action; the training stops when the total number of training rounds is reached; In the feedforward calculation of the training model, the Gaussian noise of action selection is eliminated, the optimal action is selected according to the current state space information, and the action selection is stopped when the task completion identifier value changes, thus completing the assembly task; The method for setting the pre-planned path points is as follows: Set the step length of each step of the robot arm. The starting point is the center point of the end of the robot arm gripper, and the end point is the center point of the bottom of the assembly slot. The starting and end points will change randomly during each training. According to the random assembly starting position and fixed assembly end position of each initialization scene, calculate the deviation in the x and y directions, and calculate the number of steps n required to align the center point through the deviation and step length. Then, in the first n steps, move the assembly object to align the center point of the handle with the assembly end position in the x and y directions. Keep the x and y directions of the output action always at 0, and z at -1 until it is assembled in place; The step of obtaining the current state space information of each step as the input of the autonomous assembly network for training includes: Initialize the critic1, critic2, and actor network parameters, including the weights of each fully connected layer, the number of connected layers, and the bias value; Initialize three target networks and copy the network parameters of critic1, critic2, and actor to the target networks; Initialize the experience replay pool; Set the Gaussian noise distribution range for action selection and start the loop iteration of network update; select the optimal action according to the current network model and execute it, obtain the next state space information and the current reward score, and store the current state space information, the next state space information, the selected action, the reward score, and whether the task is completed as the experience of this step into the experience replay pool; The loss function calculation errors of target network 1, target network 2, critic1, and critic2 are calculated through the Bellman equation, and the network parameters of critic1 and critic2 are updated; The actor network uses gradient ascent to update network parameters after a delay; When the total number of training rounds reaches 6000, the loop is exited and the training ends; The actor network is the execution network, and the critic1 and critic2 networks are the evaluation networks. The execution network inputs the state quantity at the current moment and maps the action taken according to the state quantity; the evaluation network inputs the state quantity at the current moment and the action taken, maps the Q value according to these quantities, and evaluates the performance of the execution network by the size of the Q value; the actor, critic1, and critic2 together constitute an autonomous assembly network; The calculation of the reward score is based on the reward function, and the reward function R is designed as follows: R＝r1+r2+r3， in, x1 and x2 are the coordinates of the center point of the assembly object handle and the coordinates of the assembly end point respectively; r2 = -max(f x , f y , f z ), f x 、f y 、f z are the three-dimensional force feedback values ​​at the end of the robotic arm; l1. They are the coordinate values ​​of the center point of the handle of the assembly object and the coordinate value of the assembly starting point; step represents the current step number of each round; The method of adding Gaussian noise with different means to the probability of selecting an action by the autonomous assembly strategy model is: When the total number of training steps reaches 2000 rounds, Gaussian noise with a mean of 0 and a variance of 0.1 is added; when the total number of training steps reaches 3000 rounds, Gaussian noise with a mean of 0 and a variance of 0.3 is added; when the total number of training steps reaches 6000 rounds, Gaussian noise with a mean of 0 and a variance of 0.5 is added.

2. The method of autonomous assembly of a robot arm suitable for an unstructured environment according to claim 1, characterized in that: The setting of the total number of training rounds includes: the total number of training rounds is 6000 rounds, of which the number of random rounds is 1000 rounds, the number of learning rounds is 5000 rounds, and the number of training steps in each round is 100 steps.

3. The method of autonomous assembly of a robot arm suitable for an unstructured environment according to claim 1, characterized in that: The state space information includes: the three-dimensional coordinates of the end of the robot arm, the three-dimensional position of the center point of the assembly object handle, the three-dimensional force feedback value of the end of the robot arm, and the coordinate value of the assembly end point.

4. The method of autonomous assembly of a robot arm in an unstructured environment according to claim 1, characterized in that: The method of selecting the optimal action according to the current network model is as follows: The network model calculates the value at the current moment based on the input state space information, and calculates the action value with the highest probability at the current moment through a greedy algorithm.

5. The method of autonomous assembly of a robot arm suitable for an unstructured environment according to claim 1, characterized in that: The movement of the end of the robotic arm is achieved by solving the inverse kinematics of the robotic arm; through a set of three-dimensional coordinate values ​​in a Cartesian coordinate system, the six-degree-of-freedom robotic arm joint angles are solved in the order of joint 1, joint 3, joint 2, joint 4, joint 5, and joint 6; the movement mode of the end of the robotic arm is to add the absolute coordinate value of the end position of the robotic arm at the current moment plus the relative motion coordinate value predicted by the autonomous assembly network as the next step of the end coordinate value of the robotic arm.

6. The method for autonomous assembly of a robot arm in an unstructured environment according to claim 1, characterized in that: When the end of the robotic arm moves along the pre-planned path point, the range of motion of the end of the robotic arm is limited: The absolute coordinate value of the end position of the robot arm at the current moment is added to the relative motion coordinate value predicted by the autonomous assembly network. When the end position of the robot arm exceeds the threshold in the x, y, and z directions, the end of the robot arm maintains the position of the previous moment.

7. A robotic arm autonomous assembly system suitable for unstructured environments, characterized in that: include: Robotic arm, three-dimensional force sensor and control system; The three-dimensional force sensor collects the forces in the x, y, and z directions received by the end of the robot arm during the assembly process when the robot arm is grasping the assembly object, and inputs the force information into the autonomous assembly network on the control system; the autonomous assembly network is constructed using the autonomous assembly method according to any one of claims 1 to 6; The autonomous assembly network predicts the next position of the robot end according to the learned assembly strategy, and the control system sends the end position to the robot drive; The robot arm performs actions according to the position sent by the control system.

Images