Datasets:
Stack bowls — LeRobot v2.1 with native depth
Instruction: Stack the bowls on the plate.
This public export contains all 81 recorded episodes (26,546 frames), with the original train/eval separation. It uses the LeRobot v2.1 episode Parquet / per-camera MP4 layout. The download revision is v2.1.
| Split | Episodes | Frames | Episode range (end exclusive) |
|---|---|---|---|
| train | 70 | 22,468 | 0:70 |
| eval | 11 | 4,078 | 70:81 |
Cameras and fields
Four cameras: cam_high, cam_left_wrist, cam_right_wrist, cam_third_person.
- RGB:
observation.images.<camera>, native 640 × 480, H.264 CRF18 MP4, dataset timeline 30 FPS. RGB uses lossy video encoding; it is not a lossless RGB export. - Depth:
observation.depth.<camera>, native uint16 millimetres, encoded losslessly as PNG bytes inside each episode Parquet. Depth is not an RGB video and is not divided by 255. Invalid zeros and the entire uint16 range are preserved. Source D2C metadata describescolor_camera_z; zero_is_invalid. observation.state,observation.velocity,observation.effort,action: float32 vectors of length 14, ordered left 6 joints + gripper, then right 6 joints + gripper. Recorded values are copied without rescaling or replacing the gripper values. Fieldactionfollows the source converter's action alignment; seemeta/recipe.json.base_action: float32 vector of length 2.observation.camera_timestamps.<camera>: original float64 capture timestamps. The regulartimestampis the dataset's index / 30 timeline, not a claim that camera arrival intervals were perfectly uniform.- Standard frame, episode, global and task indices are retained.
meta/conversion.json and per-episode receipts retain the source split/file mapping, capture metadata, camera intrinsics, D2C calibration, and synchronization settings. These are source calibration values; this export does not infer camera-to-robot extrinsics or TCP poses. Original-size images avoid changing the source intrinsics.
Read RGB and depth
The repository includes a frozen copy of the official LeRobot v0.3.3 source (dataset format v2.1), plus a small depth reader adapter in tools/. Use a separate Python environment from LeRobot v3. Dependencies are listed in tools/requirements-reader.txt; PyTorch and torchvision must be a compatible pair.
from pathlib import Path
import sys
from huggingface_hub import snapshot_download
repo_id = "Kensag/Stack_bowls_lerobot_v21"
root = Path(snapshot_download(repo_id=repo_id, repo_type="dataset", revision="v2.1"))
sys.path.insert(0, str(root / "tools"))
from read_lerobot_v21_depth import DepthLeRobotDataset
dataset = DepthLeRobotDataset(repo_id, root=root, split="train", video_backend="pyav")
sample = dataset[0]
rgb = sample["observation.images.cam_third_person"] # float32 [3,480,640], 0..1
depth_mm = sample["observation.depth.cam_third_person"] # float32 [1,480,640], raw mm
valid = depth_mm > 0
depth_m = depth_mm / 1000.0
Use split="eval" for held-out episodes or split=None to explicitly read all episodes. The adapter defaults to training episodes and selects statistics for the selected episodes. RGB transforms apply only to RGB. The old torchvision uint16-to-signed-int16 conversion is avoided; no depth values are normalized implicitly.
Parquet-only loaders expose numeric fields and embedded depth; the LeRobot reader resolves RGB from the MP4 files using episode indices and timestamps. They do not need extra copies of the RGB images in Parquet.
Validation and statistics
Before publication, every saved depth pixel was checked against source-depth hashes, state/action/capture timestamps were compared, all video frame counts were checked, and the official v2.1 SDK plus adapter were exercised on train/eval samples. See meta/validation.json and meta/reader_validation.json for the actual reports. The uint16 decoder/adapter was separately checked across all 65,536 possible values.
meta/stats.json follows the standard all-episode metadata layout. For training, use meta/stats_train.json or the provided reader's training-only statistics, not held-out episodes. RGB statistics sample up to 32 frames per episode; numeric and depth statistics use all frames, including invalid zero-depth pixels. Filter invalid depth explicitly for geometry.
The source HDF5 recordings and earlier v3 export were not overwritten. Uploads exclude transient conversion locks, local caches, and ffmpeg logs. The vendored LeRobot code retains its upstream license; no additional dataset license is asserted here.
- Downloads last month
- 432