SimFactoryLitLampSingle-v0

Memory-dependent whole-body manipulation demonstrations for the Unitree G1, in the same LeRobot v2.1 layout as SIMPLE's G1WholebodyTabletopGraspMP-v0.

Task. Watch the two lamps on the table while you load the apples; one of them lights up. Load both apples into the container, hug the container, and set it down on the pad in front of the lamp that lit up.

The lamp is lit only briefly, so the policy must remember which side lit up until the set-down.

Contents

Item	Value
Episodes	90 (50 left lamp, 40 right lamp)
Frame rate	50 fps
Camera	observation.images.egocentric (G1 head_stereo_left), 640x360, h264
states / action	32-d / 36-d, same semantics as SIMPLE (see meta/modality.json)
Format	LeRobot codebase v2.1

Generation

• Physics: MuJoCo, scripted whole-body expert (SIMPLE G1 controller).
• Rendering: NVIDIA Isaac Sim replay of the recorded trajectories.
• Domain randomization per episode, SIMPLE style: lighting, table, ground and object materials, and 1 to 3 render-only GraspNet distractors.
• Curation: only complete episodes with the container placed within 5 cm of the target pad.

Known limitation

The lamp cue is 3 s long. An automatic check finds the lit lamp in the egocentric view at the moment it lights in 49 of 90 episodes (left 43/50, right 6/40). In the others the lamp lights while it is at the edge of, or outside, the camera view and may only come into view later in the cue or not at all. Per-episode results are in meta/cue_check.json.

Files

• data/, meta/, videos/: the LeRobot v2.1 dataset.
• simple/SimFactoryLitLampSingle-v0.zip: the same files as one archive, in the layout of USC-PSI-Lab/psi-data (simple/<task>.zip). Ψ₀ training tools download this archive at a pinned revision.
• meta/stats_psi0.json: the normalization statistics Ψ₀ uses, identical to meta/stats.json.

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("randing2000/SimFactoryLitLampSingle-v0")