Energy Approximated Dynamic Subattractor for Adjusting Obstacle Avoidance Trajectories.

Area of Science:

• Robotics
• Machine Learning
• Control Systems

Background:

• Imitation learning facilitates human-robot skill transfer but struggles with environmental variations.
• Ensuring learned skills remain effective across different environments is a significant challenge.

Purpose of the Study:

• To propose a stable autonomous dynamic system (DS) for robust human-robot skill transfer.
• To enhance trajectory accuracy and disturbance resistance in learned robotic skills.

Main Methods:

• Introduced an energy-approximated dynamic subattractor (EADA) method for disturbance resistance.
• Dynamically selected subattractors using Neum, an energy function from demonstration data.
• Combined EADA with velocity modulation algorithms for global stability and precise control.

Main Results:

• Achieved global stability, precise obstacle avoidance, and autonomous trajectory recovery.
• Demonstrated effective handling of complex scenarios, including multiple and dynamic obstacles.
• Validated through simulations and real-world experiments on single-arm and dual-arm robots.

Conclusions:

• The proposed framework enables robots to learn and execute skills robustly in diverse and challenging environments.
• The EADA method significantly improves disturbance resistance and generalization capabilities in imitation learning.
• The approach ensures smooth, accurate obstacle avoidance trajectories, enhancing practical human-robot interaction.