pravsels/arx5-robot-dataset
ARX5 Manipulation Dataset Robotic manipulation demos collected on a single ARX5 arm, of tabletop tasks with dual RGB cameras and DINOv2 embeddings of the images. Converted from Ville Kuosmanen's LeRobot v2.1 datasets into a unified HDF5 format. Hardware Robot: ARX5 single arm (7-DoF) Cameras: Wrist-mounted + front-facing scene camera Stats Trajectories 1,704 Total frames 860,076 Frames per trajectory 4 – 3,034 (median 397) Total… See the full description on the dataset page: https://huggingface.co/datasets/pravsels/arx5-robot-dataset.
ARX5 Manipulation Dataset
Robotic manipulation demos collected on a single ARX5 arm, of tabletop tasks with dual RGB cameras and DINOv2 embeddings of the images.
Converted from Ville Kuosmanen's LeRobot v2.1 datasets into a unified HDF5 format.
Hardware
- Robot: ARX5 single arm (7-DoF)
- Cameras: Wrist-mounted + front-facing scene camera
Stats
Quick Start
import h5py
with h5py.File("arx5_datasets_part_001.h5", "r") as f:
traj = f["trajectory_0042"]
actions = traj["actions"][:]
images = traj["camera_0"][:]The dataset is split across 18 HDF5 files (100 trajectories each), from arx5_datasets_part_001.h5 to arx5_datasets_part_018.h5
Each part contains trajectories renumbered trajectory_0000 through trajectory_0099 within the chunk.
What's Inside
Tasks: Picking coffee props, dropping a footbag into a dice tower, rice cooker assembly (multi-step), plus DAgger correction runs.
Per frame:
actions/states— 7-DoF joint commands and observationscamera_0/camera_1— 480×640 RGB (wrist / scene)cam_rs_embd/cam_zed_embd— DINOv2 patch embeddings (256×384)
Source Datasets
This dataset aggregates the following HuggingFace datasets:
License
Apache 2.0