Precise and dexterous robotic manipulation via human-in-the-loop reinforcement learning.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Robotic manipulation faces challenges with traditional methods requiring extensive manual design and data collection.
• Limitations in reliability, speed, and robustness hinder real-world deployment of current robotic manipulation techniques.
• Reinforcement learning (RL) offers autonomous skill acquisition but struggles with sample efficiency and safety in real-world applications.

Purpose of the Study:

• To develop a human-in-the-loop, vision-based reinforcement learning (RL) system for dexterous robotic manipulation.
• To overcome limitations of existing methods by enabling autonomous skill acquisition with improved sample efficiency and safety.
• To demonstrate the system's capability on a diverse range of complex manipulation tasks under realistic industrial conditions.

Main Methods:

• Integration of human demonstrations and corrections with sample-efficient RL algorithms.
• Direct learning of RL policies in the real world using a vision-based system.
• System-level design incorporating precise assembly, dynamic manipulation, and dual-arm coordination tasks.

Main Results:

• Achieved strong performance across various dexterous manipulation tasks, including precise assembly and dual-arm coordination.
• Outperformed baseline methods by 2x in task success and executed tasks 1.8x faster on average.
• Demonstrated near-perfect success rates within 1 to 2.5 hours of real-world training.

Conclusions:

• Reinforcement learning (RL) can effectively learn complex, vision-based robotic manipulation policies directly in the real world within practical training times.
• The human-in-the-loop approach significantly enhances sample efficiency and performance for real-world robotic manipulation.
• This work paves the way for advanced learned robotic manipulation techniques in industrial applications and research.