PhysCaP: Grounding Code-as-Policy Agent with Physics-Informed Exploration
Researchers presented PhysCaP, a Physics-Informed Code-as-Policy agent designed for active perception during robotic manipulation tasks. The system adds a physics-informed exploration layer to existing code-as-policy frameworks. This enables agents to move beyond passive observation of demonstrations and instead probe objects to infer latent physical properties such as mass distribution or friction.
Current vision-language-action policies primarily imitate recorded demonstrations. They often lack mechanisms to gather information about physical characteristics that are not directly visible. PhysCaP targets this gap by grounding code generation in physics-informed exploration steps.
The approach extends code-as-policy methods with an explicit layer for physics-based information gathering. No performance metrics or deployment details were provided in the announcement.