mlnha/vace-aug-224-gr00t

VACE-augmented RoboCasa dataset — 224 episodes (GR00T / LeRobot v2.1)

The exact 224-episode training set used to fine-tune GR00T-1.5 in the VACE-augmentation baseline. Built by swapping a target object into RoboCasa pick-and-place episodes with the VACE video-diffusion model, then converting to the GR00T "gr00t_views" (LeRobot v2.1) format.

• 224 episodes, 62,445 frames, 3 camera views (left_view / right_view / wrist_view = RoboCasa robot0_agentview_left / robot0_agentview_right / robot0_eye_in_hand), 20 fps.
• 8 target objects (32 each; teapot_7 mostly excluded by a runtime cap): objaverse donut_5, steak_8, teapot_6, teapot_7; lightwheel Jar023, MeasuringCup009, SoapDispenser010, SyrupBottle005.
• Data config single_panda_gripper: state 16-d (base pos 3 + base quat 4 + EE pos rel 3
  • EE quat rel 4 + gripper qpos 2), action 12-d (base motion 4 + control mode 1 + EE pos 3
  • EE rot 3 + gripper close 1), language in annotation.human.task_description.

VACE repaints pixels only; the state/action rows are copied unchanged from the source RoboCasa episodes (verified LeRobot action order — gripper at dim 11).

Layout

data/chunk-000/episode_NNNNNN.parquet
videos/chunk-000/observation.images.robot0_agentview_left/episode_NNNNNN.mp4
                 .../robot0_agentview_right/  .../robot0_eye_in_hand/
meta/{info,modality,episodes,episodes_stats,tasks,stats}.json / .jsonl
extras/   (eval helpers; training never reads it)

Load

from gr00t.data.dataset import LeRobotSingleDataset
from gr00t.experiment.data_config import DATA_CONFIG_MAP
from gr00t.data.schema import EmbodimentTag
dc = DATA_CONFIG_MAP["single_panda_gripper"]
ds = LeRobotSingleDataset("path/to/this/dataset", dc.modality_config(),
                          transforms=dc.transform(),
                          embodiment_tag=EmbodimentTag("new_embodiment"),
                          video_backend="decord")

Trained checkpoints from this dataset: mlnha/vace-batch64-30k-ckpts.