Datasets:
so101-ttt-sim2real-controls
Control inputs for cagataydev/so101-ttt-sim2real-cosmos3: 10,000 MuJoCo SO-101 tic-tac-toe episodes rendered as depth, segmentation and edge maps, plus the expert's joint labels. 20.45 GB, 80,023 files.
episode_NNNNN/ (NNNNN = source_episode in the cosmos3 meta/flywheel.jsonl):
| file | content |
|---|---|
depth_{top,wrist}.mp4 |
8-bit linear depth, top 0.15-0.70 m, wrist 0.02-0.60 m |
seg_{top,wrist}.mp4 |
fixed-palette segmentation (sky, table, sheet, card, bar, x, o, robot) |
edge_{top,wrist}.mp4 |
Canny edges of the sim RGB |
record.parquet |
LeRobot v3 rows: action, observation.state (6 joints, degrees; gripper 0-100 %), timestamp, frame_index |
episode.json |
task, seed, n_frames, prompt, prompt_index |
All video h264, 640x480, 30 fps; the cosmos3 dataset keeps every second frame (15 fps). MANIFEST.jsonl has one row per episode. tools/ holds the generation scripts and prompt sets v2-v4 used on Hugging Face Jobs. variants/palette-v2/ is a 4-episode palette test.
Raw sim RGB: cagataydev/so101-ttt-sim2real-sim-rgb. License Apache-2.0.
- Downloads last month
- 131