MSRA, UTS and Tsinghua unveil UniSteer: injecting human corrections into RL for flow-matching VLAs
jiqizhixin · x · 2026-09-11
Microsoft Research Asia, University of Technology Sydney, and Tsinghua University present UniSteer, tackling the interface gap between human guidance and lightweight RL for flow-matching VLA policies.
- Real-robot RL is too expensive for blind exploration, so human guidance is the natural fix — but flow-matching VLAs optimize initial noise, and human actions have no direct mapping into that noise space.
- UniSteer bridges this via approximate action-to-noise inversion: when a human takes over and corrects the robot, the correction is inverted into target noise.
- The data flows into two buffers simultaneously: a demonstration buffer for direct MSE supervision of the actor, and an RL buffer that helps the critic learn these high-value state-noise pairs. Both supervised and reinforcement learning update the same lightweight noise actor while the pretrained VLA backbone stays frozen.
More from Embodied
- OpenTac tactile sensing project to be showcased at CoRL — _krishna_murthy · 2026-10-28
- Hybrid LLM agent scores 83.2%, beats nine VLA policies on new SO-101 robot arm benchmark — Truman Hickok · 2026-10-28
- cixliv: robot fighting may be the first realized market where bots beat average humans — cixliv · 2026-10-28
- Can the Hiwonder MentorPi T1 run fully offline as a robot pet? Pi 5 questioned — oulu2006 · 2026-10-28
- Testing Mint Modeler to Build the Menlo Research Robot — freelerobot · 2026-10-28
- Khosla-backed startup builds a 'mind-reading' cap with 100,000 tiny chips to write AI prompts — Distinct-Question-16 · 2026-10-28