Stanford's FACT: Force-Aware Robot Foundation Model Trained on 2,000h Human Data
yining_hong · x · 2026-10-29
Yining Hong and collaborators (with Li Fei-Fei, Yejin Choi, Jiajun Wu; first author Zhi Cao) released FACT, a Force–Action–Track multimodal robot foundation model (arXiv:2610.19155). Key points:
- Builds EgoForce, a 2,000-hour force-rich human manipulation dataset, filling the gap that most human demonstration data lacks force measurements.
- Two-stage architecture: a Track-Guided Coarse Planner jointly predicts actions, future video, and visual tracks for cross-embodiment transfer; a Force-Action Joint Refiner predicts upcoming forces and refines coarse actions.
- Evaluated on 10 dexterous tasks requiring force feedback — tool use, musical instruments (the demo video shows the robot playing a harp duet), and precision wet-lab procedures — adapting with as few as 10 robot trajectories plus 90 human ones.
More from Embodied
- University of Virginia's robot at International Lunabotics competition looks impressively smooth — TinfoilTricorn · 2026-10-29
- Neuralink's first speech restoration data: ALS participant hits 75 WPM in week one — shivon · 2026-10-29
- MOSS, a mostly 3D-printed litter-picking rover, hits 100 GitHub stars; kits coming — chrismatthieu · 2026-10-29
- Odyssey-3 world model simulates general environments in real time with greater physical accuracy — Scobleizer · 2026-10-29
- One year after 1X opened $20,000 NEO humanoid robot pre-orders, 2026 delivery still in doubt — ChrisGPT · 2026-10-29
- Asimov's open-source humanoid robot is making incredible progress in the open — jonstephens85 · 2026-10-29