Robotics researchers debate tactile prediction with vision-pretrained models
KyleMorgenstein · x · 2026-10-16
Robotics researchers Kyle Morgenstein and XOS9000 debate whether tactile prediction transfers easily to video-pretrained models.
- XOS9000 argues training a multimodal diffusion policy from scratch on the same tactile data would likely work just as well — if DVA resembles DreamerZero, adding tactile prediction and masking vision reduces it to a policy with action pretraining.
- Morgenstein pushes back: the scale is different, small models aren't strongly aligned for video prediction the way DVA is; tactile is nominally OOD for a video-pretrained model, actions only enter via the IDM, and doing a no-vision task with a vision-based model is genuinely hard.
A substantive exchange on cross-modal (tactile/vision) pretraining transfer in embodied AI.
More from Embodied
- Indie dev's AI PASSPORT wearable firmware hits 1,200+ downloads in days — ezshine · 2026-10-16
- 1X 开设加州工厂，批量量产 NEO 人形机器人 — emmanuelvivier · 2026-10-16
- Waymo 无人驾驶出租车对圣安东尼奥等三城全面开放 — emmanuelvivier · 2026-10-16
- Ropedia teases Super Week: 6 embodied AI releases in 5 days from world models to bimanual control — liuziwei7 · 2026-10-16
- A Hong Kong grocery store is actually a robot data collection lab inside Auki's office — LoganGrasby · 2026-10-16
- Peking University's SimForcing distills simulation motion priors into robot world models — PekingUniversity · 2026-10-16