NUS Proposes LIT to Break Robot Vision-Action Shortcuts

NUS introduced Latent Interface Training (LIT), which trains pose-conditioned action priors before injecting visual information via a pose-supervised latent interface, addressing vision-action shortcut learning in robot foundation models and improving LIBERO-Plus success rates by up to 10.7 percentage points.

2026-09-14 ~ 2026-09-14 · 2 related posts

• NUS releases LIT to break vision-action shortcuts in robot foundation models — NationalUniversityofSingapore · 2026-09-14
• Latent Interface Training breaks vision-action shortcuts, lifts robot success by up to 16.7 points — DJiafei · 2026-09-14