CoRL 2026 paper steers frozen flow-matching VLAs with spatial cues, nearly doubling task success
zhoubolei · x · 2026-11-09
Bolei Zhou's group presents "Cue the Flow" (CoRL 2026): a grounding module converts free-form delivery instructions into spatial cues, and a lightweight contrastive cue-conditioned adapter steers a frozen flow-matching VLA by applying a diagonal affine transform over generated action chunks — no policy fine-tuning needed. Across tabletop and mobile-base setups, it improves instruction following on in- and out-of-domain objects, with up to 2× average task success at negligible inference overhead. Authors from UIC and UCLA.
More from Embodied
- Robotics data in 2027: mono is dead, stereo scarce, tactile is the new frontier — Scobleizer · 2026-11-09
- Rhoyn demos YAM Ultra teleoperated robot arms in gentle massage mode ahead of CoRL 2026 — Scobleizer · 2026-11-09
- Unitree G1 + Mac mini demo reportedly hides 200 H100s behind it — yongqianme · 2026-11-09
- Ropedia W0 trains robot world models on 10,000 hours of human video, lifting LIBERO-Plus by 6.76 points — jiqizhixin · 2026-11-09
- StackChan robot now cheers along when its owner throws a celebratory fist pump — alexcovo_eth · 2026-11-09
- UW team's Success-Guided Sampling scales RL to 1M simulated robots for contact-rich manipulation — KyleMorgenstein · 2026-11-09