SteerAct lifts robot action-model success ~10% with just 24 rollouts per task
pliang279 · x · 2026-11-13
SteerAct is a few-shot temporal steering framework that improves pretrained robot action models at test time, spotlighted at CoRL's Learn@Deploy Workshop.
- Problem: improving a pretrained action model from only a few rollouts is hard; activation steering is lightweight but requires deciding which features to modify and when/how strongly.
- Method: a pretrained sparse autoencoder supplies candidate features, and a shared temporal descriptor summarizes each feature's distribution of local activation-order patterns, enabling cross-rollout comparison without timestep alignment. Offline, one feature per task is selected whose descriptor banks best separate successful and failed rollouts, with reference banks retained; online, the same descriptor guides intervention timing and steering coefficient.
- Results: using only 24 outcome-labeled rollouts per task, task success improves roughly 10% over unsteered models across π0.5, Fast-WAM, RoboTwin 2.0, DexJoCo and real-world manipulation, with frozen action-model weights.
- Code and demos are released.
More from Embodied
- Qwen shows in-context learning of body-part steering, sharing a map across bodies — Sauers_ · 2026-11-13
- Gemini Robotics 2.5 Debuts at CoRL 2026: Humanoid Ties a Bow Autonomously — ZoubinGhahrama1 · 2026-11-13
- LLM edits 65 embedding rows to fix robot failures, lifting held-out success 74.3% to 87.5% — 机器之心 · 2026-11-13
- Astra vs Fable 5.2 zero-shot excavator test: only visual reference, one model prevails — Scobleizer · 2026-11-13
- Waymo rider pulls up next to a Cybercab — with a human driver still in the seat — kscottz · 2026-11-13
- SimpleVLN: native multimodal LLMs navigate continuously without a specialized backbone — jiqizhixin · 2026-11-13