This paper proposes BiHapticVLA, an extension to HapticVLA for bimanual manipulation with explicit inter-hand force coordination. It introduces a Dual-Arm Safety-Aware Reward-Weighted Flow Matching objective, an Inter-Hand Force Coordination module, and a Bimanual Tactile Distillation framework. The model aims to enable safe, contact-rich bimanual manipulation without requiring tactile sensors at deployment.
Key findings
BiHapticVLA introduces three key innovations for bimanual manipulation.
The DA-SA-RWFM objective learns coordinated force-aware policies from varying tactile quality demonstrations.
The IHFC module models force exchanges as a dynamic constraint graph using graph neural networks.
The BTD framework transfers force coordination capabilities to a vision-proprioception-only student model.
Limitations & open questions
Challenges identified include asymmetric force distribution, synchronization timing, and cross-modal alignment.
Mitigation strategies for the identified challenges are proposed.