Stanford's FIDO robot dog learns physics by playing fetch with humans
kevin_zakka · x · 2026-11-12
Stanford and Toyota Research Institute present FIDO, a quadruped robot that collects in-the-wild dynamics data by playing fetch: it observes a thrown ball's trajectory, retrieves it, and returns it, creating a naturally engaging, repeatable interaction that yields grounded observations of object dynamics.
- A VLM dynamics model fine-tuned on this data transfers directly to unseen sports prediction and balloon-catching tasks, and supports robot planning.
- The authors argue interaction diversity from embodied play is a key driver of learning generalizable physical dynamics, sidestepping the limits of internet data and infrastructure-heavy collection setups.
Paper and project page are public.
More from Embodied
- Scaling H2R workshop at CoRL 2026 probes scaling laws for human-to-robot learning — mangahomanga · 2026-11-12
- EqC paper: introspective robot policies that know when to ask for help — du_yilun · 2026-11-12
- iPhone Duo fold test hits record 413,000 cycles in 13 days with zero failures — SumitGup · 2026-11-12
- kevin_zakka adds tactile sensing support to mjlab robotics framework — kevin_zakka · 2026-11-12
- MCP Display turns an old Android phone into a programmable screen for AI agents — PatrickCavalcanti · 2026-11-12
- Dennis Hong showcases THOR hazard-response robot and ARTEMIS at AIVEX Frontier 2026 — DennisHongRobot · 2026-11-12