Method for learning robot object grasping poses, server for learning robot object grasping poses, and system for learning robot object grasping poses

6-DOF Object Grasping Pose Learning System Combining Deep Reinforcement Learning and Gram-Schmidt Orthogonalization

This technology converts spatial information acquired through an input device into point cloud data and applies a deep reinforcement learning model to infer the Q-value (success rate) and rotation vector for each point. Based on the inferred values, it uses Gram-Schmidt orthonormalization to determine the 6-DOF grasping pose.

Conventional supervised learning-based grasping techniques require sophisticated dynamic models, limiting their use for grasping unknown objects without CAD data. Furthermore, the presence of multiple objects often leads to occlusion, which reduces grasping success rates.

This technology uses a deep reinforcement learning model to infer grasping positions and angles directly from point clouds without the need for label generation. In the event of a learning failure, it calibrates the reward function model through inverse reinforcement learning and optimizes the 6-DOF grasping pose using the Gram-Schmidt orthogonalization technique. Applicable to logistics picking, service robots, and manufacturing automation, it significantly increases the grasping success rate for unknown objects even without CAD data.

‍

‍

Key Features:

• Deep reinforcement learning model training phase, which learns from point cloud data and infers the Q-value and rotation vector for each point.
• Work area spatial information extraction phase, where the data generation unit extracts the portion corresponding to the robot device's workspace from the first spatial information to create second spatial information.
• Point cloud data generation phase, which includes a spatial information reception phase where the learning server's data generation unit receives the first spatial information.
• Spatial information acquisition phase, which acquires the first spatial information containing the target object for grasping through an input device.

‍

This invention was developed with the support of the Artificial Intelligence Convergence Innovation Talent Cultivation program by the Ministry of Science and ICT.