Generalizable Multistage Assembly via One-Shot Category-Level Demonstration.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Imitation learning enables robots to acquire skills from demonstrations.
• Current methods require numerous demonstrations, leading to time-consuming data collection.
• Existing approaches struggle with generalization to novel scenarios and object variations.

Purpose of the Study:

• To develop a generalized multistage manipulation network for category-level robot assembly tasks.
• To enable robots to learn from a single demonstration and generalize to new object instances.
• To address performance degradation in novel scenarios and during unexpected collisions.

Main Methods:

• Utilized category-level pose estimation to extract manipulation trajectories from demonstrations.
• Applied manipulation-pose generalization to transfer trajectories to novel object instances.
• Implemented real-time action correction using force feedback for adaptive control.

Main Results:

• The proposed network successfully learned a multistage screw-nut assembly task from a single demonstration.
• The system generalized effectively to new object instances with varying shapes and sizes.
• Real-time action correction allowed adaptation to unexpected collisions during task execution.

Conclusions:

• The generalized multistage manipulation network offers an effective and flexible solution for robot skill acquisition.
• This method significantly reduces data collection requirements and enhances generalization capabilities.
• The approach demonstrates robustness in handling novel scenarios and dynamic environmental changes.