SteerAct lifts real-robot success from 31.1% to 52.2% by steering one feature
liuzhuang1234 · x · 2026-10-21
A Stanford × MIT × Harvard team unveiled SteerAct: with a frozen VLA/WAM model, only 24 test-time rollouts per task and steering a single feature, real-robot mean success jumps from 31.1% to 52.2% (+21.1 pp, 68% relative), with no weight updates. Demo video available.
Related event: SteerAct Steers VLA Inner Features at Test Time, Boosting Robot Success(2 posts)→
More from Embodied
- Robotics startup launches Deployments to serve fine-tuned HF checkpoints to robots in real time — arnie_hacker · 2026-10-21
- Amazon FAR's FIBER uses action-aligned latent planning to let Unitree G1 open doors and carry boxes — ChongZzZhang · 2026-10-21
- Sentdex trains robots to go from lying down to a half kneel, eyeing a full stand — Sentdex · 2026-10-21
- NVIDIA open-sources MotionBricks: 350K motion skills at 15,000 FPS, 2ms latency — aftahi_ai · 2026-10-21
- NVIDIA Cosmos3-H-Dreams: Generative Simulation Brings Clinical Surgery to Life — Hugging Face Blog · 2026-10-21
- FLoTS-SLAM uses optical-flow residuals to hit 0.73 DM-IoU in dynamic monocular SLAM — zhenjun_zhao · 2026-10-21