Reproduce “UniSteer: Unified Noise Steering for Efficient Human-Guided VLA Adaptation” 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
Junjie Lu, Xinyao Qin, Yuhua Jiang, Kaixin Wang, Chuheng Zhang, Bin Liang, Jun Yang, Min Xu and 1 more
Researchers from the University of Technology Sydney and Tsinghua University, in collaboration with Microsoft Research, developed UniSteer, a framework that efficiently adapts diffusion-based Vision-Language-Action (VLA) models by translating human action-space corrections into noise-space supervision. The method achieved an average task success rate of 90% across diverse real-world robotic manipulation tasks, up from 20% initially, and demonstrated 100% out-of-distribution generalization in multiple scenarios with fewer interactions.
Key insights
- The noise-to-action mapping of flow-matching VLA policies is bijective, enabling a unique noise variable to exist for any given action, which is key for inversion.
- A recursive fixed-point iteration effectively and efficiently approximates the inverse of each Euler step, allowing action-space human corrections to be converted into noise-space targets.
- Integrating human-guided noise-space supervision with online reinforcement learning significantly accelerates policy adaptation and improves exploration efficiency compared to methods relying solely on autonomous exploration or action-space imitation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.