sim-pnp-multitask-prior-n10-libero

A 1,694-episode LeRobot v2.1 dataset of 13 single-object pick-and-place tasks in OmniGibson / BEHAVIOR-1K, collected automatically using a cuRobo motion-planning shortcut seeded from a real teleop grasp prior. Same 3-camera LIBERO schema we use for π0.5 sim SFT.

This is the auto-collected sibling of IDEAS-Lab-Northwestern/sentinel-pnp-multitask (human teleop) — same task family, methodologically distinct (scripted phase-A grasp + closed-loop place).

Quick facts

Codebase version	LeRobot v2.1
Robot	Franka Panda (longfinger)
Total episodes	1,694
Total frames	2,505,357
FPS	30
Tasks	13 (single-object pnp, "middle of the table" → "green goal")
Cameras	3 × 256×256 H.264 (image_left, image_right, wrist_image)
State	8D (EEF xyz + axis-angle rxyz + gripper L/R)
Actions	7D EEF-delta (dpos_xyz, drot_xyz) + gripper
Chunks	2 (chunk-000: eps 0–999, chunk-001: eps 1000–1693)
Size on disk	~7.7 GB

Collection methodology

Collected on a vast.ai dual-RTX 4090 instance over 18.5 h per GPU (37 GPU-hours total), driven by the phase_a_grasp_from_dataset shortcut in SENTINEL-Lite:

1. Phase A — grasp: an example successful grasp is read out of a small real-teleop LeRobot prior dataset for each (task, seed) cell. cuRobo plans and executes the approach + close in OmniGibson.
2. Phase B — place: closed-loop OSC EEF-delta controller (action_normalize=False, assisted grasping) drives the object to the green goal region.
3. Per (task, seed) cell the pipeline keeps running until 10 successful episodes are recorded (the "n10" in the dataset name). Cells with persistent failures are skipped.
4. The two GPUs ran disjoint (task, seed) shards; their outputs were merged afterwards (episode indices and the global index column reindexed, chunks_size=1000 re-shuffled).

Driver logs and the merge script are preserved alongside the source captures on the collection box; the per-shard targets were 99 (task, seed) units done + 15 skipped per GPU (114 planned).

Tasks

task_index	Prompt
0	pick up the teacup in the middle of the table and place it at the green goal
1	pick up the mug in the middle of the table and place it at the green goal
2	pick up the coffee cup in the middle of the table and place it at the green goal
3	pick up the goblet in the middle of the table and place it at the green goal
4	pick up the bowl in the middle of the table and place it at the green goal
5	pick up the cocktail glass in the middle of the table and place it at the green goal
6	pick up the chalice in the middle of the table and place it at the green goal
7	pick up the decanter in the middle of the table and place it at the green goal
8	pick up the saucepan in the middle of the table and place it at the green goal
9	pick up the water glass in the middle of the table and place it at the green goal
10	pick up the gravy boat in the middle of the table and place it at the green goal
11	pick up the beaker in the middle of the table and place it at the green goal
12	pick up the beer glass in the middle of the table and place it at the green goal

File layout (LeRobot v2.1)

data/chunk-000/episode_000000.parquet … episode_000999.parquet
data/chunk-001/episode_001000.parquet … episode_001693.parquet
videos/chunk-{000,001}/{image_left,image_right,wrist_image}/episode_NNNNNN.mp4
meta/info.json
meta/tasks.jsonl
meta/episodes.jsonl
meta/episodes_stats.jsonl

Loading

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("IDEAS-Lab-Northwestern/sim-pnp-multitask-prior-n10-libero")
print(len(ds), ds.features)
sample = ds[0]  # frame 0 of episode 0

Schema details

state  : float32 [8]  = (eef_x, eef_y, eef_z, axisangle_x, axisangle_y, axisangle_z, gripper_l, gripper_r)
actions: float32 [7]  = (dpos_x, dpos_y, dpos_z, drot_x, drot_y, drot_z, gripper)
image_left   : video 256×256×3 RGB, h264, 30fps
image_right  : video 256×256×3 RGB, h264, 30fps
wrist_image  : video 256×256×3 RGB, h264, 30fps

Intended use

LoRA fine-tuning of vision-language-action policies (π0.5, OpenVLA-style) on diverse single-object pnp prompts where the grasp prior is scripted by cuRobo MP rather than learned from a small set of human teleops. Pairs well with the human-teleop counterpart for ablations on data-collection methodology.

Sister datasets (same org)

• sentinel-pnp-multitask — human teleop counterpart
• sim-stack-flat-30-libero, sim-stack-same-30-libero — same schema, stack family
• sim-dusty-transfer-30-libero, sim-lid-transport-food-30-libero — same schema, other task families

License

Apache-2.0. The underlying BEHAVIOR-1K assets are subject to their own terms of use.