Stanford's PsiBorg: one 8B next-token-prediction model acts as policy, simulator, tracker
chris_j_paxton · x · 2026-11-10
A Stanford team is demoing PsiBorg at CoRL: an 8B universally promptable robot model for open-world manipulation, trained from scratch fully in-house.
Key idea: treat policy, world model, and value learner (temporal distance) all as prompts, trained with just next token prediction — a surprisingly strong pretraining objective. One model simultaneously works as a policy, a simulator, and a progress tracker, and generalizes zero-shot to new objects and goal images. Explicitly not a VLA or WAM. Demos at CoRL 2026, JW Marriott Austin, Lone Star (3rd floor).
More from Embodied
- Arduino open-sources RouterBridge_ROS: real-time MCU I/O talks directly to ROS 2 — kscottz · 2026-11-10
- Agibot humanoid robot walks hand-in-hand through crowded conference floor — chris_j_paxton · 2026-11-10
- TRI's Haruki Nishimura to talk trustworthy deployment of large behavior models — GlenBerseth · 2026-11-10
- First CyberCab rides impress: 'I think it's exceeded Waymo in experience' — carlosdponx · 2026-11-10
- NVIDIA unveils GR00T 2 world-action model: 40% zero-shot success, human demos transfer to robots — yukez · 2026-11-10
- AI is the enabler, manufacturing is the moat: diode on AI-accelerated hardware design — espricewright · 2026-11-10