PhysCaP: NTU and NVIDIA's agents measure mass and stiffness with no extra sensors
shaohua0116 · x · 2026-11-13
PhysCaP, from National Taiwan University, NVIDIA Research and NYCU, is a physics-informed code-as-policy agent system accepted at NeurIPS 2026's Robot Learning Workshop and CoRL 2026's Agentic Robotics Workshop, with code released.
Key points:
- VLA policies rely on passive observation and can't infer latent properties like mass or stiffness; PhysCaP adds a physics-informed exploration layer that seeks information through interaction.
- Training-free modules estimate mass (lift 15cm, read joint torques; 38.1g error) and stiffness (squeeze mapped to 5 levels, e.g. avocado ripeness) from proprioception alone — no extra sensors.
- A Planner decides when to explore and stop; a Prioritizer filters implausible interactions via heuristic scoring to balance cost and information gain.
- On real tabletop tasks (identify empty can, pick ripe avocado, pack groceries) plus LIBERO simulation, it reaches full success with fewer interactions than baselines and works with Gemini, GPT and Claude VLMs.
More from Embodied
- BoSS benchmark released: 180K+ stellarators, 3TB of simulation data for generative models — ZhongingAlong · 2026-11-13
- Fly brain hooked up to a synth learns to sing Bad Apple — Bedrovelsen · 2026-11-13
- CoRL takeaways: SuperMimic steals the show, dexterity and tactile sensing improve, human data still a bottleneck — KyleMorgenstein · 2026-11-13
- Outworld Labs launches simulation platform with new physics engine for robot learning — ycombinator · 2026-11-13
- YC-backed Mantle launches AI-native maintenance service for robots with 250+ technicians — ycombinator · 2026-11-13
- UMI diffusion policy hits 100/100 cube-stacking success with slowed demos — DominiqueCAPaul · 2026-11-13