Reproduce “SUREFlow: State-space Uncertainty-aware REsidual Flow Matching for Robust Robot Manipulation” on OpenResearch

About this paper

Md Tanvir Islam, Sai Navaneet Peddapalli, Sangmoon Lee, Sangtae Ahn

SUREFlow, a framework from Kyungpook National University, introduces a meta-reinforcement learning approach that integrates a Mamba-based state-space model with uncertainty-aware residual flow matching for robust robot manipulation. This method achieves an average success rate of 92.5% on the LIBERO benchmark, outperforming existing transformer and Mamba-based policies, and attains long-horizon robustness comparable to models with significantly more parameters (7 billion vs. 179.1 million).

Key insights

• Modeling heteroscedastic uncertainty in the flow velocity field allows the policy to quantify prediction confidence for each action dimension, differentiating reliable from unreliable predictions.
• Internal action refinement based on predicted uncertainty can effectively mitigate compounding integration errors and enhance stability during long-horizon rollouts without relying on online environment feedback.
• Leveraging state-space models like Mamba provides a computationally efficient backbone that achieves competitive robustness with significantly fewer parameters compared to large transformer-based VLA models.

Summary by alphaXiv.