PhysiCalWorld Calibrates World Models to Scene Physics with Label-Free In-Context Interaction Clips
Scobleizer · x · 2026-11-03
PhysiCalWorld tackles world model drift caused by hidden physics properties (mass, friction, joint damping) invisible to visual observation. An embodied agent first collects an exploratory interaction clip in the test scene, which is encoded into a compact physics descriptor token that grounds future predictions in the scene's actual dynamics.
- No fine-tuning, no analytical physics modeling, no physics labels — trained end-to-end with latent prediction loss alone
- Descriptor extracted once at inference, reused across rollouts at negligible cost
- Extrapolates beyond physics ranges seen during training
Paper, code and project page are public.
More from Embodied
- Swedish grocer ICA hit by cyberattack that remotely set smart freezers to 50°C — josh_bickett · 2026-11-03
- Industrial designer uses fal to iterate real hardware designs from photos, faster prototyping — OdinLovis · 2026-11-03
- NVIDIA Cosmos Simulation Recipes Turn Daytime Drives Into Rainy Nights With Trajectories Intact — NVIDIAAI · 2026-11-03
- AI Agent Performs Brain Drug Injection on a Live Mouse — ChuckDBrooks · 2026-11-03
- 3D World Simulator predicts appearance, geometry and touch at 50+ FPS on an RTX 5090 — YunzhuLiYZ · 2026-11-03
- Google's Suncatcher satellite runs Gemma 270mi n space—power limits mean only one TPU lit at a time — demian_ai · 2026-11-03