Dex-One2Many: Learning Dexterous Manipulation from a Single Human Demonstration

TL;DR

Dex-One2Many is a framework that learns dexterous manipulation from a single human video, enabling generalizable policies. It uses sequential scene graphs to guide reinforcement learning, allowing for efficient exploration. This approach achieves zero-shot transfer to a real multi-fingered hand.

Why It Matters

Dex-One2Many addresses the challenges of generalization and exploration in dexterous manipulation, making it a promising approach for real-world applications. Its ability to learn from a single human video reduces the need for costly robot demonstrations. This has significant implications for robotics and automation.

Key Findings

• Dex-One2Many exceeds baselines by 6.5% in seen configurations
• it widens the gap to 71% in unseen scenarios
• it achieves zero-shot transfer to a real multi-fingered hand