Reproduce “ForceFlow: Learning to Feel and Act via Contact-Driven Flow Matching” 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
Shuoheng Zhang, Yifu Yuan, Hongyao Tang, Yan Zheng, Qiaojun Yu, Pengyi Li, Guowei Huang, Helong Huang and 2 more
ForceFlow learns to perform contact-rich robotic manipulation by integrating temporal force history and active force prediction with a Flow Matching backbone and a hierarchical Vision-to-Force handover. The framework significantly enhances task success and force regulation, exhibiting robust generalization across varied physical properties and spatial configurations.
Key insights
- Decoupling spatial reasoning from physical interaction through the V2F handover mechanism is critical for achieving robust generalization to both spatial and physical out-of-distribution scenarios.
- Asymmetric multimodal fusion, which injects force signals globally into the network via Adaptive Layer Normalization, effectively prevents higher-dimensional visual features from masking crucial force information.
- Jointly predicting future contact forces alongside motion commands enables the policy to learn the underlying force-motion coupling, leading to proactive force regulation and more compliant robot interactions.
- The use of temporal force history significantly enhances stability by filtering sensor noise and providing context for contact detection, while active force prediction refines interaction quality and smoothness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.