DreamFAR world model trained from scratch on robot data only, ~$110k in H200 hours
k7agar · x · 2026-10-30
- DreamFAR is a world model trained from scratch using only robot data (no internet pretraining), supporting real-time teleoperation, minutes-long rollouts, and generalization across 1,000+ tasks.
- A commenter notes the training cost: 36,864 H200 GPU hours, roughly $110k at market rates.
- It shows capable robot world models don't necessarily require massive web-scale pretraining.
More from Embodied
- SpAItial launches Echo-3: minutes of phone capture become real-scale 3D Gaussian Splat worlds — rmbrualla · 2026-10-30
- Amazon FAR's DreamFAR: a world model trained from scratch on robot data only, 1,000+ tasks — _krishna_murthy · 2026-10-30
- Tesla FSD demo avoids a stop-sign runner, wins over an Elon-averse mom considering Rivian — BenCostaTesla · 2026-10-30
- XPENG's IRON humanoid robot walks off its new production line — ingliguori · 2026-10-30
- PokeBot's RoboAPI brings zero-shot, natural-language robot control to any compatible arm — haider1 · 2026-10-30
- Tesla Robotaxi hosts Accessibility Day as feedback flows into Cybercab and future products — EricETesla · 2026-10-30