PhysCaP Teaches Code Policies to Probe Hidden Mass and Stiffness Before Grasping

Passive vision hits a wall

Pick and place demos look impressive until the target object is hidden, empty, or ripe versus firm. Standard vision language action policies imitate what cameras show, yet many manipulation failures trace to latent physics: mass, stiffness, contents. PhysCaP, described on arxiv.org/abs/2608.21031, adds an exploration layer on top of code as policy frameworks so robots can gather that information deliberately.

Training free property modules

The agent ships training free modules that estimate mass and stiffness from proprioception during small exploratory pushes and lifts. No extra force torque sensor is required beyond what many arms already expose. A Planner decides when exploration is worth the time cost and when the policy should commit to a grasp. A Prioritizer filters implausible interactions and ranks the remainder with a heuristic score so the robot does not thrash randomly.

Tasks and results

Authors evaluate real tabletop scenarios such as finding hidden objects, detecting empty cans, and selecting ripe avocados, plus a simulated LIBERO style task. Passive baselines fail when properties are occluded. Naive interactive baselines over explore and waste time. PhysCaP matches or beats their success rates with fewer interactions and shorter execution time, according to the preprint.

Ablation studies credit the mass and stiffness extractors individually: removing either module raises failure rates on tasks where that property dominates.

Why code policies

Using executable policies rather than end to end pixels to torques makes the exploration logic inspectable. Safety engineers can audit the Planner's stop rules and cap the force envelope for a probe motion. That matters as warehouses begin piloting humanoids in partially structured cells where every unplanned shove has a liability line item.

Sources

arxiv.org/abs/2608.21031 PhysCaP preprint, August 2026<br />Project page physcap.github.io