Reproduce “VFP: Variational Flow-Matching Policy for Multi-Modal Robot 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
Xuanran Zhai, Qianyou Zhao, Qiaojun Yu, Ce Hao
Researchers from the National University of Singapore and Shanghai Jiao Tong University developed the Variational Flow-Matching Policy (VFP), an imitation learning framework that overcomes the mode-averaging problem in flow-matching models while maintaining fast inference. VFP enables robots to generate distinct, multi-modal actions for complex manipulation tasks, demonstrating improved success rates in both simulated and real-world experiments.
Key insights
- Explicitly conditioning a flow-matching policy on a learned latent variable can effectively decompose velocity ambiguity, allowing it to capture and separate distinct behavioral modes.
- Applying Optimal Transport regularization is crucial for ensuring that the learned policy's action distribution faithfully covers all modes of the expert data, thereby mitigating mode averaging.
- A Mixture-of-Experts architecture provides the necessary capacity to represent complex multi-modal policies efficiently by sparsely activating relevant experts for specific behavioral modes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.