behavior_r1_pro_val

A held-out validation split of the BEHAVIOR-1K R1Pro simulation data, packaged as a single merged LeRobot v2.1 dataset.

Contents

• 250 episodes — the last 5 episodes (indices 195–199) of each of the 50 sim_behavior_r1_pro.task-* tasks, merged and renumbered 0…249.
• 3,490,956 frames, fps 30, robot R1Pro.
• 50 tasks (one language instruction each).
• 3 camera streams per episode: head, left_wrist, right_wrist (all 256×256 h264).

This is the exact hold-out set excluded from training in our Cosmos3 GR-1/BEHAVIOR world-model experiments (the built-in heldout_per_task=5 rule: episodes with episode_index >= total_episodes - 5 per task).

Layout

meta/
  info.json          # total_episodes=250, total_frames=3490956, total_tasks=50, splits={"train":"0:250"}
  episodes.jsonl     # 250 rows, episode_index 0..249
  tasks.jsonl        # 50 tasks, task_index 0..49
  stats.json         # recomputed over the 250 episodes
  modality.json
  embodiment.json
data/chunk-000/       episode_00000000.parquet … episode_00000249.parquet
videos/chunk-000/<camera>/episode_00000000.mp4 … episode_00000249.mp4

Per-frame parquet features include action (23-DOF joint), observation.state (256-D proprio), observation.cam_rel_poses (21), observation.task_info (46), timestamp, and annotation.human.coarse_action. The global index, per-episode episode_index, and task_index columns were rewritten for the merged dataset; all other columns are copied verbatim from the source.

Source & attribution

Derived from nvidia/PhysicalAI-Robotics-GR00T-X-Embodiment-Sim (the sim_behavior_r1_pro.* task folders), built on the BEHAVIOR-1K benchmark. This repo is a re-packaged subset for evaluation convenience; all original data ownership, terms, and licensing follow the upstream NVIDIA dataset and BEHAVIOR-1K — please consult those sources before use.