What it is
Reasoning-annotated robot navigation data combining video, actions, instructions, and frame-level CoT across three benchmarks for dynamic-environment VLA training. 3 robot navigation benchmarks with raw trajectories plus 20-second segmented episodes linking RGB sequences, synchronized actions, language instructions, frame-level CoT, history, and future trajectory supervision.
Key results
- 3 source benchmarks: SCAND, GND, and DynaNav
- 20-second overlapping navigation episodes
- CC-BY-4.0 license
Useful for training and evaluating robot navigation models that predict actions from video, instructions, and temporal history.