PhysCaP: Grounding Code-as-Policy Agent with Physics-Informed Exploration
The agent actively tests objects to infer hidden physical properties before choosing a manipulation policy.
PhysCaP adds a physics-informed exploration layer to code-as-policy robotics, estimating mass and stiffness from robot proprioception without extra sensors. A Planner decides whether more probing is needed, while a Prioritizer ranks plausible interactions to avoid wasteful exploration. The paper reports real tabletop tests on hidden-object search, empty-can detection, and ripe-avocado finding, plus a LIBERO simulation. Against passive and naive interactive baselines, PhysCaP matched performance with fewer interactions and shorter execution time. HF Daily Papers' note
PhysCaP adds a physics-informed exploration layer to code-as-policy robotics, estimating mass and stiffness from robot proprioception without extra sensors. A Planner decides whether more probing is needed, while a Prioritizer ranks plausible interactions to avoid wasteful exploration. The paper reports real tabletop tests on hidden-object search, empty-can detection, and ripe-avocado finding, plus a LIBERO simulation. Against passive and naive interactive baselines, PhysCaP matched performance with fewer interactions and shorter execution time. HF Daily Papers' note
score 5