Notes from the lab.

MC-PhysBench

We give video world models one frame of a simple physics scene, ask them to predict what happens next, measure how long each prediction stays physically valid, and document what causes failure.

Optimizing Argus for Action Labelling

Raising Argus from 29.8% to 50.2% semantic F1 on WGO-Bench with gpt-6.1-sol-high.

Introducing MC-EgoHands

From first-person recordings to structured 3D motion. Review, retarget and export human demonstrations into robot-learning workflows.