LIT lifts real-robot success rates 13-17 points by breaking vision-action shortcuts
DJiafei · x · 2026-09-11
- Paper from NUS Magic Lab et al.: VLAs and WAMs learn vision-action shortcuts (lighting, background, camera view correlated with actions), hurting OOD generalization.
- LIT trains in two stages around a shared spatial goal: first train the action expert without images (language, robot state, target end-effector pose), then route vision through latent tokens supervised to predict the same pose.
- Real-robot results on a YAM dual-arm platform: in-distribution 88.0% vs 74.7% baseline (+13.3pp); new lighting +16.7pp, new camera +16.7pp, distractors +13.3pp. Only LIT redirects when the instructed goal changes.
- Sim results on LIBERO-Plus improve average success across four architectures including π0.5. Paper, code, and project page are open.
Related event: LIT Breaks Vision-Action Shortcuts in Robot VLA Models(4 posts)→
More from Embodied
- NVIDIA's ThinkWorld learns robot skills from unlabeled human videos via chain-of-dynamics reasoning — nvidia · 2026-10-28
- Just Action Model: decoupling what separates VLA from world-action models — Yonghao Yu · 2026-10-28
- NVIDIA Research releases Foundation4D: unified 4D perception model turns camera video into 3D scenes, no lidar needed — Scobleizer · 2026-10-28
- Dyna Robotics visits four hotel laundry rooms before starting Taku POC — JasonMa2020 · 2026-10-28
- Dyna engineer on Taku: from 24-hour napkin folding to full laundry workflows — JasonMa2020 · 2026-10-28
- Apple spent 20 years trashing touch-screen laptops — now it's making one — Scobleizer · 2026-10-28