Stanford debuts PsiBorg at CoRL: one 8B model serves as policy, simulator, and progress tracker
dyamins · x · 2026-11-10
Stanford researchers demo PsiBorg at CoRL 2026: an 8B universally promptable robot model for open-world manipulation, trained from scratch fully in-house — not a VLA or world action model.
PsiBorg frames robotics capabilities as any-to-any token prediction, letting one model act simultaneously as a policy, a simulator, and a progress tracker. Pretraining mixes internet videos of people/scenes/objects, robot videos across embodiments, and paired observation-action data.
Key finding: first predicting the temporal distance Δt to the goal and folding it into the goal image substantially improves action chunk prediction versus conditioning on the goal image alone, enabling generalization to unseen objects and diverse goals.
Demo runs Nov 9 (12:00–18:30) and Nov 10 (10:00–18:30) CST at Demo #2, Lone Star, JW Marriott Austin. Paper, code, and model weights are all marked SOON.
More from Embodied
- Fable 5.2 posts 64.9% on RoboDojo robot control, beating most purpose-built systems — claudeai · 2026-11-10
- 5.5M-param tactile encoder hits real hands, lifts pi0.5 scores from 48 to ~70 — Alternative_Set4042 · 2026-11-10
- Pincer1, a six-legged robot hardware wallet, opens preorders, fully open source — R4nsack · 2026-11-10
- FlipAnything at CoRL: robots learn complex behaviors from LLM-generated code — berkeley_ai · 2026-11-10
- BAIR brings 33 main-conference papers to CoRL 2026 in Austin — berkeley_ai · 2026-11-10
- The Second Station: why 5x-ing one factory station doesn't save the line — saranormous · 2026-11-10