SeonghuJeon/leftright-p1-corrected

Leftright unified benchmark data

Simulated bimanual YAM demonstrations in RoboTwin 2.0 for left-arm (ego) policies that cooperate with an uncontrolled right-arm partner. Eleven tasks; each scene has a no-pause episode and a partner-pause episode (1 s or 2 s). Episodes are stored in the native Epishelf BRCD format (one dataset per episode, mkv/ffv1:gop=1 lossless video, Arrow tables), exactly as used by the training code.

config	status	episode rows	payload dirs	splits
main	completed	704	704	train 572, validation 132
role_ablation	completed	704	704	train 572, validation 132

role_ablation is the mirrored (opposite-role) view, excluded from main training. It is partial: Hanging Mug mirrored episodes are not collected yet (640 of 704 references). Its 128 Lift Pot / Grab Roller rows are symmetric-role aliases that point to the main payload directories, so downloading role_ablation also needs those episodes/main/... folders (the download command fetches them).

Layout

• episodes/<config>/<episode_id>/epishelf/ native BRCD dataset: RobotDataset.open(path) yields one episode with streams head, wrist_left (RGB uint8), proprio [N,14], action [N,1,14], action_valid, physics_step, timestamp_ns, language, offline_annotations, camera parameters, end poses, pointcloud.
• episodes/<config>/<episode_id>/{metadata,storage,subgoals}.json: collection metadata and frame labels; labels.jsonl is the subtask-label file the S1/S2 preparation reads.
• manifests/<config>.json: the full training manifest (per-episode task, split, kind, pause_seconds, scene_seed, frames, role fields, prompt, invalid frames, verification); paths relative to the release root.
• index/<config>/<split>.jsonl: flat per-episode index for browsing.
• release.json, checksums.sha256 (sha256sum -c checksums.sha256).

Use

from ego2world_research.datasets.hf_release import download
manifest = download("<org>/<repo>", "data/leftright-hf", "main", output="data/leftright-hf-main.json")
from ego2world_research.integrations.robotwin.monolithic import EpishelfDataset
train = EpishelfDataset(manifest, "train", horizon=32)

Without the repo: from epishelf import RobotDataset; RobotDataset.open("episodes/main/<id>/epishelf")[0] (Epishelf commit c36c3dd65bca3a1ddc6646cc400e7c5954029cd1). Observations advance every 15 physics ticks (250 Hz physics, 16.7 Hz frames); action is the next observed joint sample, not an executed controller command.

Licenses (unresolved; do not assume)

No license has been chosen for this dataset; license: other is a placeholder until the owner decides. Components: RoboTwin 2.0 code is MIT; RoboTwin 2.0 object/texture assets (Hub TianxingChen/RoboTwin2.0) were not license-checked here. YAM kinematics (i2rt-robotics/i2rt, MuJoCo Menagerie i2rt_yam) are MIT; the YAM appearance URDF/meshes come from NVIDIA Isaac Omniverse content under terms not verified here. Fabrica assembly parts (yunshengtian/Fabrica) are MIT. Rendered images depict these assets.