RoboCasa365 pretrain atomic: human demos + GR00T rollouts (LeRobot v2.1)

Per-task LeRobot v2.1 datasets (loadable with lerobot==0.3.3, the version pinned by robocasa) that merge, for each of 17 RoboCasa365 atomic tasks (pretrain split):

• the official human teleop demos (pretrain/atomic/<Task> from robocasa v1.0), and
• policy rollouts from six GR00T N1.5 variants (gr00tFMAtomicPosttrain10, gr00tFMAtomicPosttrain100, gr00tFMAtomicPosttrain30, gr00tFMAtomicTargetOnly100, gr00tFMPretrainAll, gr00tMTPretrainHuman), each with action chunk length Ta=16, 30 rollouts each, run at the robocasa v1.0.1 per-task horizon with ignore_done=True.

Built for off-policy evaluation (FQE); total 11.4 GB.

Tasks

Task	succ / total	rollout succ / total	task horizon	human len	frames	GB
CloseBlenderLid	124 / 286 (43%)	18 / 180 (10%)	900	237–637	198,933	0.86
CloseFridge	221 / 285 (78%)	115 / 179 (64%)	900	100–522	187,988	0.83
CloseToasterOvenDoor	244 / 289 (84%)	134 / 179 (75%)	450	115–291	100,365	0.45
CoffeeSetupMug	166 / 285 (58%)	61 / 180 (34%)	600	160–354	131,636	0.61
OpenCabinet	192 / 286 (67%)	85 / 179 (47%)	1050	169–618	225,442	1.11
OpenDrawer	208 / 282 (74%)	106 / 180 (59%)	750	135–250	155,488	0.64
OpenStandMixerHead	242 / 286 (85%)	133 / 177 (75%)	450	80–358	93,061	0.47
PickPlaceCounterToCabinet	194 / 288 (67%)	86 / 180 (48%)	750	161–309	159,225	0.78
PickPlaceCounterToStove	220 / 286 (77%)	112 / 178 (63%)	600	145–386	130,839	0.65
PickPlaceDrawerToCounter	178 / 283 (63%)	75 / 180 (42%)	750	200–489	166,819	0.84
PickPlaceSinkToCounter	232 / 288 (81%)	124 / 180 (69%)	900	171–368	188,397	0.86
PickPlaceToasterToCounter	217 / 284 (76%)	112 / 179 (63%)	600	170–412	134,307	0.63
SlideDishwasherRack	203 / 277 (73%)	103 / 177 (58%)	450	126–288	98,702	0.61
TurnOffStove	159 / 287 (55%)	50 / 178 (28%)	750	157–490	166,241	0.68
TurnOnElectricKettle	223 / 287 (78%)	115 / 179 (64%)	450	72–191	93,010	0.40
TurnOnMicrowave	155 / 287 (54%)	48 / 180 (27%)	450	97–195	95,010	0.43
TurnOnSinkFaucet	225 / 286 (79%)	118 / 179 (66%)	600	140–394	131,195	0.56

Human demos end 16 frames after success is first held (robocasa collect_demos.py), so their length varies. Rollouts always run the full task horizon; success is read from next.success per frame and is_success per episode, not from episode end.

Layout

pretrain_atomic/ mirrors robocasa's pretrain/atomic/ download; one LeRobot v2.1 dataset per task.

pretrain_atomic/<Task>/lerobot/
  data/chunk-000/episode_XXXXXX.parquet
  videos/chunk-000/observation.images.<cam>/episode_XXXXXX.mp4   3 cams, 256x256 h264, 20 fps
  meta/info.json  meta/stats.json  meta/modality.json  meta/embodiment.json
  meta/tasks.jsonl  meta/episodes.jsonl  meta/episodes_stats.jsonl
  meta/fqe_masks.json         episode-index lists: train valid success fail human rollout <policy>
  extras/dataset_meta.json    env_args + provenance
  extras/episode_XXXXXX/      states.npz, model.xml.gz, ep_meta.json, rewards_dense.npz

Episodes are ordered human first, then rollouts. meta/episodes.jsonl rows carry source (human|rollout), policy, is_success, split (train|valid), layout_id, style_id, source_episode_index.

Features

Identical to the official robocasa365 LeRobot export, plus two columns:

• observation.state float64[16]: base_pos(3) + base_quat(4) + eef_pos_rel(3) + eef_quat_rel(4) + gripper_qpos(2)
• action float64[12]: base_motion(4) + control_mode(1) + eef_delta_pos(3) + eef_delta_aa(3) + gripper(1); binary dims in {-1, 1}
• observation.images.robot0_agentview_left / _right / _eye_in_hand: video 256x256x3
• next.reward float32: sparse success r(s'); next.done bool: true at the final frame only
• next.reward_dense float32 (extra): dense task-progress reward Φ(s') replayed from sim states
• next.success bool (extra): per-frame env._check_success()
• annotation.human.task_description, annotation.human.task_name, timestamp, frame_index, episode_index, index, task_index

Loading

hf download infope/robocasa --repo-type dataset --include "pretrain_atomic/CloseBlenderLid/*" --local-dir ./robocasa

import json
from lerobot.datasets.lerobot_dataset import LeRobotDataset  # lerobot==0.3.3

root = "./robocasa/pretrain_atomic/CloseBlenderLid/lerobot"
ds = LeRobotDataset(repo_id="CloseBlenderLid", root=root)
masks = json.load(open(f"{root}/meta/fqe_masks.json"))  # e.g. masks["rollout"], masks["success"]