Bi-DexHands: Towards Human-Level Bimanual Dexterous Manipulation.

Area of Science:

• Robotics
• Artificial Intelligence
• Cognitive Science

Background:

• Human-level dexterity in robotics is an unsolved challenge, particularly for tasks requiring bimanual cooperation.
• Existing reinforcement learning (RL) environments are limited for complex, coordinated robotic hand manipulation.
• Simulators often lack the complexity to train robots for human-like motor skills.

Purpose of the Study:

• To introduce Bi-DexHands, a new simulation environment for dexterous bimanual robotic manipulation.
• To evaluate various reinforcement learning algorithms on bimanual manipulation tasks.
• To assess the performance of RL in single-agent, multi-agent, offline, multi-task, and meta-RL settings.

Main Methods:

• Developed Bi-DexHands in Isaac Gym, featuring two dexterous hands, 20 bimanual tasks, and thousands of objects.
• Achieved high-throughput RL training (over 30,000 FPS on an NVIDIA RTX 3090).
• Conducted comprehensive evaluations of popular RL algorithms across different learning paradigms.

Main Results:

• On-policy algorithms like PPO successfully learned simple manipulation tasks, comparable to a 48-month-old's motor skills.
• Multi-agent RL enhanced performance in tasks requiring skilled bimanual cooperation (e.g., lifting, stacking).
• Current RL algorithms struggle with learning multiple manipulation skills in multi-task and few-shot scenarios.

Conclusions:

• Bi-DexHands provides an efficient platform for advancing bimanual robotic dexterity research.
• While progress is made in simpler tasks, complex multi-skill manipulation remains a significant challenge for RL.
• Further research is needed to improve RL's capabilities in multi-task and few-shot bimanual manipulation.