Reproduce “FWBC-VLA: Force-Aware Whole-Body Compensation for Contact-Rich Loco-Manipulation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yutian Zhang, Siyuan Ma, Liwen Yang, Yang Li, Ce Hao, Haozhen Chi, Dong Wei, Qiaojun Yu and 1 more
FWBC-VLA enables a wheeled-legged robot to open doors and wipe boards despite contact disturbances. It estimates interaction from joint-torque and proprioceptive signals, without a dedicated force/torque sensor, and uses the estimate for task actions and bounded body compensation. In whiteboard wiping, final-stage success was 64% versus ForceVLA’s 24% over 25 trials; this is one stage, not a universal margin.
Key insights
- A dual-head LSTM architecture (HSR-Force), combining history and state experts with a fixed gate, effectively estimates interaction forces from proprioceptive data, outperforming single-expert or simpler methods.
- Integrating sensorless interaction feedback into VLA policies via "force tokens" significantly enhances manipulation actions, while a separate, parallel whole-body compensation module provides robust stability for the robot's base under sustained contact.
- A hierarchical control strategy, where interaction representations simultaneously adapt manipulation actions and mitigate body-level disturbances, is crucial for robust performance in contact-rich loco-manipulation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.