Datasets:

qualiadev
/

qsim-shelf-restocking-mujoco-300

episodes int64	success int64	failed int64	placements int64	steps int64	frames_per_camera int64	duration_s float64	bytes int64	arm_counts dict	shelf_counts dict	carton_counts dict	unique_ordered_plans int64	seed_min int64	seed_max int64	checks list
300	300	0	600	542,518	108,640	10,850.36	9,684,681,808	{ "left": 297, "right": 303 }	{ "middle": 300, "top": 300 }	{ "e14": 300, "e27": 300 }	64	42	341	[ "success, released, supported, settled", "episode IDs and seeds", "finite state and action arrays", "50 Hz observation/action timing", "10 Hz aligned camera indices", "all RGB chunks decoded; 320x240 uint8", "terminal state timing", "SHA-256 for every episode" ]

Qsim MuJoCo Shelf Restocking

300 seeded synthetic shelf-restocking episodes with the stock bimanual OpenArm robot, physical gripper contacts, smooth scripted motion, and synchronized robot head and wrist camera observations.

Release status: complete — 300/300 episodes passed the task and dataset checks.

600 placements; 542,518 observation/action samples; 108,640 RGB frames per camera; 3.01 hours of simulated trajectories; 9.68 GB of HDF5 episodes.

Arm selections: {'left': 297, 'right': 303}. Shelf placements: {'middle': 300, 'top': 300}. The batch covers 64 distinct ordered placement plans.

Camera previews

The robot head view is on the left; left and right wrists are stacked on the right.

Episode 1 · Episode 2

Episode setup

Seeds 42–341, one independently reset episode per seed.
Two placements per episode: one E14 carton and one E27 carton.
Randomized carton order, left/right arm per placement, shelf assignment, and destination slot. Each episode uses both the middle and top shelf.
Four candidate slots per arm, two on each shelf row. Source and destination Y coordinates are mirrored into the chosen arm's workspace.
Cameras, lighting, colors, carton dimensions, and station geometry are fixed.
MuJoCo 3.13.0, Mink 1.2.0 IK with DAQP, stock robot meshes, primitive shelf/tote.
1,000 Hz physics, 50 Hz absolute joint-position targets, 10 Hz RGB capture.
Three RGB streams at 320 × 240: context (robot-mounted head), wrist_left, and wrist_right. The near-clipping plane is 1 mm.

The native task evaluator requires each carton to be near its target, upright, released from the gripper, supported by the intended shelf, and stable. Collection uses the existing Qsim task implementation; four isolated local worker processes accelerate the batch without changing episode sampling or physics settings.

Files and format

episode_000000.h5 through episode_000299.h5: one complete episode per file.
run.json: resolved configuration, library versions, model provenance/hashes, action semantics, and named joint order.
summary.json: episode outcomes and placement evidence.
manifest.jsonl: episode index, seed, length, camera frame count, placements, byte size, and SHA-256 checksum.
validation.json: aggregate counts and checks performed on the release.
generator.json: source hashes identifying the local generator implementation.
FORMAT.md: detailed HDF5 schema and timing conventions.
previews/: combined robot-head and wrist MP4s for selected episodes.

This is the native Qsim HDF5 format. It is not a LeRobot-format dataset. Download the files and open them with h5py; a Hugging Face tabular viewer is not required.

Observations and RGB are captured before the associated action is applied. camera_sample_index maps each RGB frame to a 50 Hz observation/action row. Actions are 16 joint-position targets in radians, including actuated gripper joints; the right gripper uses negative opening angles. Full simulator state is stored separately under privileged/.

Download and read

import json
import h5py
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="qualiadev/qsim-shelf-restocking-mujoco-300",
    repo_type="dataset",
    filename="episode_000000.h5",
)
with h5py.File(path, "r") as episode:
    metadata = json.loads(episode.attrs["metadata"])
    state = episode["observations/joint_positions"][:]
    actions = episode["actions/joint_position_targets"][:]
    head_rgb = episode["cameras/context"][0]
    camera_times = episode["time"][:][episode["camera_sample_index"][:]]

Download the full public dataset:

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="qualiadev/qsim-shelf-restocking-mujoco-300",
    repo_type="dataset",
    local_dir="qsim-shelf-restocking-mujoco-300",
)

Scope and provenance

This is an uncalibrated simulation baseline. Task success is a simulator outcome and does not establish transfer to a physical robot. Randomization is discrete; multiple seeds can select the same ordered placement plan. No train/test split is assigned. Account for repeated plans when designing an evaluation split.

The stock robot originates from enactic/openarm_mujoco, vendored via the Q-research revision recorded in run.json. Geometry hashes are included in each episode. Robot assets are not redistributed in this dataset.

The dataset is public and ungated. No additional dataset license is asserted in this release.

Downloads last month: 61

Number of rows:

Total file size:

9.69 GB