RLobot-jun/bigenlight_multitask_gr00t_30per_task

Bigenlight / Theo four-task UR7e corpus — bigenlightmultitaskgr00t30pertask

Original demonstrations: Bigenlight / Theo. GR00T conversion and subset packaging: RLobot-jun. This is a derived training-format dataset, not new data collection.

120 unique episodes · 46,660 frames · four tasks · 30 Hz. First 30 source episode indices per task (ascending, not random). All selected episodes are exposed as train; no held-out evaluation split or measured policy success rate is claimed. The 30/50/all releases overlap and must not be treated as disjoint train/test sets.

Contents and processing

• —Episode-level LeRobot v2.1 files, compatible with the GR00T loaders tested here.
• —n17/ and n15/ expose the same episodes, with separate compatible metadata. Each view contains real standalone files, not links to the publisher's server. They are NOT different splits. Download only the view you need to avoid storing the video payload twice locally.
• —Absolute UR7e joint commands: six joint positions in radians + one gripper command; state is the corresponding measured seven-dimensional vector. Gripper convention: 0=open, 1=closed. This is not EEF-delta action data.
• —Two RGB cameras: cam1 scene and cam2 wrist, 1280×720 at 30 Hz. No depth.
• —Action horizon 16 in both supplied modality configurations. Native model padding is 40×132 for N1.7 and 16×32 for N1.5; only the real 16×7 actions are used.
• —Original task text and float32 state/action values are unchanged. Episode and global frame IDs are contiguous in this release; source_mapping.json preserves original IDs and immutable upstream revisions.
• —Aggregate source AV1 videos were frame-accurately re-encoded as per-episode H.264, CRF18, at original resolution. Video compression is lossy. All clip frame counts were verified; sampled RGB alignment was checked against sources.
• —Normalization statistics were recomputed on this selected corpus only. N1.7 metadata includes its statistics fingerprints; N1.5 omits those reserved entries. N1.7 uses percentile normalization, N1.5's supplied config uses min/max.
• —No demonstrations were removed for collection/recovery tags. No reward, success or terminal labels were invented; this release is for BC, not labeled offline RL.
• —CPU loader/preprocessor verification reports are under each view's VALIDATION.json. These are data-integrity tests, not policy evaluation; GPU forward/backward and training on this release were not run for publication.

Download

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="RLobot-jun/bigenlight_multitask_gr00t_30per_task", repo_type="dataset",
    local_dir="bigenlight-gr00t",
    allow_patterns=["n17/**", "configs/**", "README.md", "LICENSE", "NOTICE",
                    "source_mapping.json", "release.json", "MANIFEST.json"],
)

For N1.5, replace n17/** with n15/**. For both versions, include both patterns. Use bigenlight-gr00t/n17 or bigenlight-gr00t/n15 as the dataset root. The Hub viewer configurations likewise select a version, not an additional dataset.

GR00T N1.7

From an installed Isaac-GR00T N1.7 checkout, use the absolute dataset path ending in /n17, --embodiment-tag NEW_EMBODIMENT, and --modality-config-path /absolute/path/bigenlight-gr00t/configs/n17_modality.py. The supplied config uses absolute joint targets for a consistent action meaning across versions; it does not reuse the older single-carrot relative-joint recipe.

DEAS GR00T N1.5

From the DEAS-Isaac-GR00T N1.5 environment, load configs/n15_data_config.py and register its UR7eDataConfig() instance in scripts.gr00t_finetune.DATA_CONFIG_MAP before invoking main(ArgsConfig(...)):

import runpy
from scripts import gr00t_finetune as bc

root = "/absolute/path/bigenlight-gr00t"
Config = runpy.run_path(root + "/configs/n15_data_config.py")["UR7eDataConfig"]
bc.DATA_CONFIG_MAP["ur7e_multitask"] = Config()
# Pass dataset_path=[root + "/n15"], data_config="ur7e_multitask",
# embodiment_tag="new_embodiment" in bc.ArgsConfig when starting BC training.

Training hyperparameters, GPU allocation and W&B account/project are deliberately not embedded in this dataset. Keep the N1.5 and N1.7 Python environments isolated.

Provenance and license

The four original cards declare Apache-2.0. See LICENSE, NOTICE, source links above, and the per-episode mapping. Collection details/recovery annotations remain in the originals, e.g. meta/source_takes.json; those tags are not success labels. Subset and conversion changes are described above. Pinned source revisions:

• —Bigenlight/carrot_in_pot_lighting_lerobot_v3: 77695b1bcc8cb615004742eb2ef1f9681863f44c
• —Bigenlight/bowl_stack_lighting_lerobot_v3: ab446a6bf6e6641a9eb494e00aa15c6a99fd5522
• —Bigenlight/bowl_stack_triple_lerobot_v3: 4494eb4bc5d662abe1ec74097304ef64b7461b56
• —Bigenlight/cube_stack_lerobot_v3: 6b1b2353556f2dd571198ef3b5a89f64ae65e80d