pepijn223/molmoact2-yam-human-video-pilot

MolmoAct2-BimanualYAM · generated human demonstration videos (pilot, 100 episodes)

Human first-person demonstration videos generated from robot episodes of allenai/MolmoAct2-BimanualYAM-Dataset (revision e9f21ae), one per annotated subtask, for Zero-WAM in-context (human-video) training.

• Episodes: 100 (3 per task, all 34 tasks). episode_mapping.json maps the local episode index to the source episode index.
• Videos: 1559 of 1578 subtask segments (19 failed) in human_videos/episode_XXXXXX/segment_XXX.mp4 (832×480, 24 fps, ~5 s) with the edited first frame segment_XXX_first_frame.png. human_videos.jsonl lists each segment's subtask, start/end time in the robot episode, prompts, models and status.
• Pipeline: lerobot-annotate --human_video.enabled=true (huggingface/lerobot#4809): Qwen3.8-27B (HF router) annotates subtasks and writes the prompts; Qwen/Qwen-Image-Edit-2509 removes the robot from the top-camera frame and adds human hands; MiniMaxAI/MiniMax-H3 (480P, 5 s) animates it, both on fal via HF Inference Providers.
• data/ holds the subtask annotations (language_persistent) for the 100 episodes; review_reel.mp4 shows one segment per task next to the robot's top-camera clip.

The videos are synthetic and not quality-filtered yet.