Robot teaching data generation method based on artificial intelligence

Abstract

The invention discloses a robot teaching data generation method based on artificial intelligence, and the method comprises the following steps: randomly extracting one or more objects from a preset 3D object model library containing a large number of diversified objects as an operation target of the cycle when each data generation cycle begins, a robot body is randomly selected from a preset robot model library, and the randomly selected object and the robot are instantiated into a simulation environment. The trajectory generated based on physical simulation and a classical motion planning algorithm naturally meets kinematics and dynamics constraints, is high in data quality, can be directly used for training, provides 3D point cloud and a 2D orthogonal projection drawing, perfectly supports current mainstream robot control models based on 2D vision and 3D point cloud, is completely carried out in a simulation environment, and has a wide application prospect. And equipment loss and potential safety hazards possibly caused by real robot training are avoided.

Description

Technical Field

[0001] This invention relates to the field of robot learning and simulation technology, and more specifically, to a method for generating robot teaching data based on artificial intelligence. Background Technology

[0002] Data-driven machine learning methods, especially deep reinforcement learning and imitation learning, have become the mainstream paradigm for training robotic agents. However, the success of these methods heavily relies on large-scale, high-quality, and diverse teaching data. Currently, obtaining real-world robot teaching data mainly depends on:

[0003] Expert teaching: This is carried out by professionals through a teach pendant or remote operation, which is inefficient, costly, and difficult to scale.

[0004] Autonomous robot exploration: This is carried out in real-world environments, but the process is slow, there are risks of equipment wear and tear and safety hazards, and the efficiency for exploring complex tasks is extremely low.

[0005] The data obtained through the above methods has inherent bottlenecks in terms of scale, diversity, and security, which severely restricts the generalization ability of robot models and makes them difficult to adapt to various changes in open environments.

[0006] Simulation environments offer a possibility for solving this problem. However, existing methods for generating simulation data are largely limited to:

[0007] Simple random sampling: Randomizes the position of an object within a limited range to generate a simple "approach-grab" action, lacking complex task logic and precise motion planning.

[0008] Script generation: Fixed scripts are written for specific tasks, which are inflexible, difficult to generate diverse data, and cannot adapt to dynamic changes in objects and the environment.

[0009] Ignoring physical constraints: The generated trajectory may not conform to the robot's kinematics or dynamics constraints, resulting in low data usability.

[0010] Therefore, there is an urgent need in this field for a systematic method that can automatically generate large quantities of high-quality, diverse, and physically feasible robot teaching data. Summary of the Invention

[0011] In view of the problems existing in the prior art, the purpose of this invention is to provide a method for generating robot teaching data based on artificial intelligence.

[0012] To address the aforementioned technical problems, the present invention adopts the following technical solution.

[0013] A method for generating robot teaching data based on artificial intelligence includes the following steps:

[0014] S1. At the beginning of each data generation cycle, one or more objects are randomly selected from a pre-set 3D object model library containing a large number of diverse objects as the operation target for this cycle.

[0015] S2. Randomly select a robot body from a pre-set robot model library;

[0016] S3. Instantiate randomly selected objects and robots into the simulation environment;

[0017] S4. Generate a stable grasping pose (P_{grasp}) in real time based on the geometric features of the current object.

[0018] S5. Generate a collision-free trajectory T based on a motion planning algorithm;

[0019] S6. When executing the planned trajectory, simultaneously record RGB images, depth maps, point clouds, 2D projections, status data, and motion data;

[0020] S7. After the trajectory is executed, the simulation engine checks whether the task was successful.

[0021] S8. Only store the complete data sequence of successful trials in the database in the standard dataset format.

[0022] As a further description of the above technical solution: the model library in S1 includes high-precision 3D scanned real object models or CAD models, different geometric shapes and differentiated physical attribute data.

[0023] As a further description of the above technical solution: the robot model library in S2 includes robot models with different configurations, degrees of freedom, sizes and end effectors.

[0024] As a further description of the above technical solution: S3 also has the function of randomizing the data of object pose, robot initial configuration, environment and viewpoint and task target.

[0025] As a further description of the above technical solution: In step S4, while capturing pose data, the planning algorithm adapts to the kinematics and end effector morphology of the currently selected specific robot.

[0026] As a further description of the above technical solution: the data tag of S6 must contain the object ID and robot ID metadata used in this cycle.

[0027] As a further description of the above technical solution: the execution criteria of S7 are whether the grasp is stable and whether the object is placed at the target position.

[0028] As a further description of the above technical solution: the complete data in S8 includes data on observation, action, reward and completion signal.

[0029] Compared with the prior art, the advantages of this invention are:

[0030] (1) This solution can generate data covering a large number of scenario configurations through a multi-dimensional randomization strategy, effectively improving the generalization and robustness of the learning model, realizing large-scale data production without human intervention 24 / 7, and greatly reducing the cost and time of data acquisition.

[0031] (2) The trajectory generated by this scheme based on physical simulation and classical motion planning algorithm naturally satisfies kinematic and dynamic constraints, has high data quality, and can be directly used for training. At the same time, it provides 3D point cloud and 2D orthogonal projection map, which perfectly supports the current mainstream robot control models based on 2D vision and 3D point cloud. It is carried out entirely in the simulation environment, avoiding the equipment damage and safety hazards that may be caused by real robot training. Attached Figure Description

[0032] Figure 1 This is a flowchart of the automated data generation engine described in this invention. Detailed Implementation

[0033] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0034] To address the inherent bottlenecks in the scale and diversity of current data acquisition, which severely restrict the generalization ability of robot models and make them difficult to adapt to various changes in open environments, Implementation Example 1 is proposed.

[0035] Please see Figure 1 In this embodiment, a method for generating robot teaching data based on artificial intelligence includes the following steps:

[0036] S1. At the beginning of each data generation cycle, one or more objects are randomly selected from a pre-set 3D object model library containing a large number of diverse objects as the operation target for this cycle.

[0037] S2. Randomly select a robot body from a pre-set robot model library;

[0038] S3. Instantiate randomly selected objects and robots into the simulation environment;

[0039] S4. Generate a stable grasping pose (P_{grasp}) in real time based on the geometric features of the current object.

[0040] S5. Generate a collision-free trajectory T based on a motion planning algorithm;

[0041] S6. When executing the planned trajectory, simultaneously record RGB images, depth maps, point clouds, 2D projections, status data, and motion data;

[0042] S7. After the trajectory is executed, the simulation engine checks whether the task was successful.

[0043] S8. Only store the complete data sequence of successful trials in the database in the standard dataset format.

[0044] The model library in S1 includes high-precision 3D scanned real object models or CAD models, different geometric shapes and differentiated physical property data;

[0045] The robot model library in S2 includes robot models with different configurations, degrees of freedom, sizes, and end effectors.

[0046] In this embodiment, the trajectory generated by physical simulation and classical motion planning algorithms naturally satisfies kinematic and dynamic constraints, has high data quality, and can be directly used for training. At the same time, it provides 3D point cloud and 2D orthogonal projection map, which perfectly supports the current mainstream robot control models based on 2D vision and 3D point cloud. It is carried out entirely in a simulation environment, avoiding the equipment damage and safety hazards that may be caused by real robot training. Thus, the device has the advantages of increasing the scale and diversity of data and improving its adaptability to various changes in open environments.

[0047] Furthermore, to address the issues of writing fixed scripts for specific tasks, which lack flexibility, struggle to generate diverse data, fail to adapt to dynamic changes in objects and environments, and suffer from insufficient security, a second implementation example is proposed:

[0048] Please see Figure 1 In S4, while capturing pose data, the planning algorithm adapts to the kinematics and end effector morphology of the currently selected robot.

[0049] The data tag in S6 must contain metadata about the object ID and robot ID used in this loop;

[0050] The execution criteria for S7 are whether the grasp is stable and whether the object is placed in the target position;

[0051] The complete data in S8 includes data on observations, actions, rewards, and completion signals.

[0052] In this embodiment, a multi-dimensional randomization strategy can generate data covering a massive number of scenario configurations, effectively improving the generalization and robustness of the learning model, realizing large-scale data production with 24 / 7 unattended operation, and greatly reducing data acquisition costs and time. Thus, the device has the advantages of improving data training security and reducing data acquisition costs and time.

[0053] The above description is merely a preferred embodiment of the present invention; however, the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concepts, should be covered within the scope of protection of the present invention.

Claims

1. A method for generating robot teaching data based on artificial intelligence, characterized in that: Includes the following steps: S1. At the beginning of each data generation cycle, one or more objects are randomly selected from a pre-set 3D object model library containing a large number of diverse objects as the operation target for this cycle. S2. Randomly select a robot body from a pre-set robot model library; S3. Instantiate randomly selected objects and robots into the simulation environment; S4. Generate a stable grasping pose in real time based on the geometric features of the current object; S5. Generate a collision-free trajectory T based on a motion planning algorithm; S6. When executing the planned trajectory, simultaneously record RGB images, depth maps, point clouds, 2D projections, status data, and motion data; S7. After the trajectory is executed, the simulation engine checks whether the task was successful. S8. Only store the complete data sequence of successful trials in the database in the standard dataset format.

2. The method for generating robot teaching data based on artificial intelligence according to claim 1, characterized in that: The model library in S1 includes high-precision 3D scanned real object models or CAD models, different geometric shapes, and differentiated physical property data.

3. The method for generating robot teaching data based on artificial intelligence according to claim 1, characterized in that: The robot model library in S2 includes robot models with different configurations, degrees of freedom, sizes, and end effectors.

4. The method for generating robot teaching data based on artificial intelligence according to claim 1, characterized in that: The S3 also has the ability to randomize data on object pose, robot initial configuration, environment and viewpoint, and task objectives.

5. The method for generating robot teaching data based on artificial intelligence according to claim 1, characterized in that: In S4, while capturing pose data, the planning algorithm adapts to the kinematics and end effector morphology of the currently selected specific robot.

6. The method for generating robot teaching data based on artificial intelligence according to claim 1, characterized in that: The data tag in S6 must contain the object ID and robot ID metadata used in this cycle.

7. The method for generating robot teaching data based on artificial intelligence according to claim 1, characterized in that: The execution criteria for S7 are whether the grasp is stable and whether the object is placed at the target position.

8. The method for generating robot teaching data based on artificial intelligence according to claim 1, characterized in that: The complete data in S8 includes data on observations, actions, rewards, and completion signals.

Images