Variable-Damping Impedance Control for Contact Tasks: A Reinforcement Learning Method Integrating HER and Importance...

Area of Science:

• Robotics
• Control Systems
• Machine Learning

Background:

• Force control is vital for robotic tasks like assembly, impacting accuracy, stability, and safety.
• Unstructured environments pose challenges due to uncertainty, oscillations, and friction, hindering control and reinforcement learning.

Purpose of the Study:

• To develop a robust deep reinforcement learning-based variable-damping impedance control method for robotic contact tasks.
• To enhance dynamic adaptability and training stability in uncertain environments.

Main Methods:

• Implemented a deep reinforcement learning-based variable-damping impedance control framework.
• Integrated parameterized Hindsight Experience Replay (TO-HER) for improved sample quality and efficiency.
• Utilized Importance Sampling (IS) with density ratio estimation to mitigate sample distribution mismatch and enhance training stability.

Main Results:

• The proposed method demonstrated superior convergence, final return, and training stability compared to baseline methods.
• Achieved higher tracking accuracy and improved dynamic response in contact scenarios.
• Showcased strong effectiveness and robustness for robotic contact control in unstructured environments.

Conclusions:

• The novel control method significantly improves robotic performance in contact-rich tasks within unstructured environments.
• The integration of TO-HER and IS effectively addresses training instability and enhances adaptability.
• This approach offers a promising solution for advanced robotic manipulation and interaction.