Ego2Act

Content & Construction: Ego2Act is a dataset comprising 110 real-world tasks across diverse domains and clutter levels. It includes 660 human-collected reference videos and 1,980 model-generated outputs for each case, requiring models to generate complete physical executions of everyday egocentric actions from an initial scene image.

Scoring & Findings: The benchmark uses Ego2ActJudge, a reference-free automated evaluator assessing task completion and physics plausibility. Analysis reveals that state-of-the-art video generation models frequently skip dependent steps or commit fine-grained physical errors during complex manipulations.