To grasp a bottle, a robot needs to know where to touch it, not just what it is. Robot trials are costly and simulations are narrow, so testing this 'physical intelligence' is hard. Everyday first-person (egocentric) video shows humans making such contacts naturally.
The authors built a large set of verified human interactions, each with an intended action and precise contact points, then asked current models to predict the contact before it happened. Most struggled, sometimes missing the object entirely.
Using those contacts as a lookup memory for a robotic grasping policy roughly doubled success in simulation and helped even more on a real robot. Measuring contact grounding can guide better robot learning.