UW RAIVN Lab brings GAM, MolmoAct2 and Reflex to CoRL: 3D grounding beats VLA baselines
LINJIEFUN · x · 2026-11-09
UW RAIVN Lab presents three robotics/3D vision works at CoRL:
- GAM (Grounded Action Model): trains policies on a pretrained 3D detector (WildDet3D) instead of VLA/WAM backbones — 55.3% avg success on RoboTin 2.0 (vs 52.0% Spatial Forcing), 47.6% under scene randomization (vs 30.4% Abot-M0), and 61% on LIBERO-PRO (vs 53% π0.5). Conditionable by language/point/box prompts and usable as a low-level controller.
- MolmoAct2: links embodied reasoning to continuous actions, with open models, data and training code.
- Reflex: onboard RGB-D whole-body control letting a humanoid catch thrown boxes with 1s reaction time.
More from Embodied
- iPad mini 8 hands-on: 1g heavier and a third the price of iPhone Air, 60Hz capped — oran_ge · 2026-11-09
- Dooder puts an AI with on-device hearing and voice entirely on the Apple Watch — breagz · 2026-11-09
- Stanford team teaches an AI agent a pool table's physics by letting it take real shots (CoRL) — drmapavone · 2026-11-09
- Furong Huang Lab brings six human-video-to-robot works to CoRL 2026 — furongh · 2026-11-09
- 3D animated teardown explains how the open-source local-AI microduck robot works — Individual-Dot5488 · 2026-11-09
- Three CMU robot papers spotlighted at CoRL 2026: memory, vision-touch steering, world models — pabbeel · 2026-11-09