hojjunekim/sam_turn_aug_uni

LEGS Task 3, long-horizon: SAM3D-AUG (200)

A training set of the paper's SAM3D-AUG (200) condition for Task 3 (long-horizon), from LEGS (code, paper). The mesh-only baseline: rendered with SAM 3D meshes instead of the splat, without color calibration. It re-renders the episodes of the same task's LEGS (200) set for the paper's appearance study, with one look per quarter of the 200 episodes: 0–49 with an apple and a box on a blue table, 50–99 with the orange and plate on a blue table, 100–149 unchanged, and 150–199 with an apple and a box on the original tables. The apple and box episodes carry the apple and box prompt.

Episodes	200
Frames	243,586 at 30 fps
Prompts	"pick the apple, turn right, and put it in the box" "pick the orange, turn right, and place it on the plate"
Camera	observation.images.ego_view, 640×480
State and action	observation.state_psi0_18 and action.psi0_18, the 18-D end-effector layout described under State and action
Format	LeRobot v2.1

Use

From a clone of the code repository:

bash scripts/train/download.sh --dataset hojjunekim/sam_turn_aug_uni
TASK=turn DATASET=hojjunekim/sam_turn_aug_uni bash scripts/train/psi0.sh   # or pi05.sh / groot.sh

The other sets of the paper are listed under Paper datasets.

License

This dataset is distributed under the SAM License, because its images show object and table meshes generated with SAM 3D Objects. If you redistribute it, you must do so under that license and include a copy of it. If you publish research that uses this dataset, the SAM License requires you to acknowledge the use of SAM Materials.

Citation

@inproceedings{kim2026legs,
  title     = {LEGS: Fine-Tuning Teleop-Free VLAs for Humanoid Loco-manipulation in an Embodied Gaussian Splatting World},
  author    = {Kim, Hojune and Chen, Timothy and Sun, Jiankai and Osterberg, Lars W. and Chen, Qianzhong and Wang, Ke and Schwager, Mac},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2026}
}