DexJoCo-X: Benchmarking Action Representations for Multi-Hand Dexterous Manipulation

TL;DR

The paper introduces DexJoCo-X, a benchmark for multi-hand dexterous manipulation, and evaluates the effectiveness of different action representations. The results show that a unified action space with embodiment-aware experts can enable one policy to control multiple hands. This approach achieves higher success rates than other methods.

Why It Matters

The development of scalable cross-embodiment learning is crucial for the practical application of dexterous hands in various tasks. This research has significant implications for robotics and automation, enabling more efficient and adaptable control of complex manipulation tasks. The findings can inform the design of more effective and generalizable control systems.

Key Findings

• Being-H0.5 achieves 47.7% mean success with function-aligned action slots
• Ego-Pi preserves pretrained action head but is ineffective for seven-hand joint training
• Shared action interface is insufficient for multi-hand learning