Video trajectories.
Actions, code & experience.

Standard and Random variants are listed separately as 54 gallery tasks, grouped into 42 base tasks for benchmark results.

All examples come from the completed, officially aggregated seed-1 evaluation. Full-evaluation metrics use all 2,100 scenes and the paper’s category weighting. Each displayed task’s metrics use all its evaluated scenes: 25 for each Generalization variant, 50 for other tasks. The 3 examples are a display subset.

3 examples per displayed task: the first 3 official records, without outcome filtering. Generalization Standard and Random variants appear separately, matching the official 54-task gallery. Paper metrics group them into 42 base tasks.

Web videos are compressed to 480×360 and retain every frame, frame rate, and duration. The official video clock is separate from the agent’s interaction timeline; execution selection does not imply timestamp alignment.