NUS releases LIT to break vision-action shortcuts in robot foundation models
NationalUniversityofSingapore · hf · 2026-09-14
NUS introduces LIT (Latent Interface Training), which improves robot action generalization by first training pose-conditioned action priors without images, then constraining visual inputs through a pose-supervised latent interface that preserves spatial goal information — breaking the vision-action shortcut in robotics foundation models.
More from Embodied
- Sharing Damiao J4310 actuator data with a bearing-mounted test rig and Codex-built GUIs — k7agar · 2026-09-14
- Musk: SpaceX option package gives Tesla Roadster ~10 rocket thrusters, maybe flight — seanmcdonaldxyz · 2026-09-14
- Zero-shot sim-to-real achieved for Asimov robot, community effort on Unitree G1 credited — freelerobot · 2026-09-14
- Watching an empty Tesla self-drive to a grocery store front still feels sci-fi — yunta_tsai · 2026-09-14
- MuJoCo's mjbatch lets you run thousands of physics sims in parallel on CPU — kevin_zakka · 2026-09-14
- Yacine to build PCBs and assemble ultra-cheap robot units by end of year, seeks buyers — yacineMTB · 2026-09-14