VivekSai07/franka_stacking_50

Franka stacking: 50 demonstrations

LeRobot v3.0 dataset with 50 demonstrations and 21,800 frames at 30 Hz.

Visualize episodes

Open episode 0 in the LeRobot visualizer. Change the episode index to browse the collection. Sign in to the visualizer to access a private repository. The episode preview on this dataset page uses the original MP4 streams.

observation.images.up (first camera shard; contains all 50 episodes):

observation.images.wrist (first camera shard; contains all 50 episodes):

observation.images.front (first camera shard; contains all 50 episodes):

Task instructions

• task 0: pick up the red cube and place it perfectly on top of the blue cube

Format and channels

• Episodes: 50; frames: 21,800; tasks: 1.

• Stored split: train episodes 0:50.

• data/: frame-level Parquet; meta/: schema, tasks, statistics, and episode/video offsets; videos/: MP4 camera streams.

• LeRobot v3.0 can store multiple episodes in a single Parquet/video shard; meta/episodes/ specifies their boundaries.

• action (8 values): panda_joint1.pos, panda_joint2.pos, panda_joint3.pos, panda_joint4.pos, panda_joint5.pos, panda_joint6.pos, panda_joint7.pos, gripper.pos

• observation.state (9 values): panda_joint1.pos, panda_joint2.pos, panda_joint3.pos, panda_joint4.pos, panda_joint5.pos, panda_joint6.pos, panda_joint7.pos, finger1.pos, finger2.pos

Cameras (width x height):

• observation.images.up: 640 x 480 RGB, av1
• observation.images.wrist: 640 x 480 RGB, av1
• observation.images.front: 640 x 480 RGB, av1

Load and download

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("VivekSai07/franka_stacking_50", revision="v3.0")

Download the complete repository, including videos and metadata:

from huggingface_hub import snapshot_download

snapshot_download(repo_id="VivekSai07/franka_stacking_50", repo_type="dataset", revision="v3.0")

For private repositories, authenticate with an account that has access. The tabular viewer configuration selects only frame-level Parquet files; tasks, episode summaries, and provenance are excluded from that table.

Split and normalization

Only a training split is declared. No independent validation/test set is supplied. If you create a held-out split, recompute normalization from training episodes only.

Provenance and preservation

The original payload was already uploaded at commit e2b3a413d726cb8dc8943768d9b6bf94679541b8 on v3.0; it was restored to the default main branch on 2026-10-09. Original collection software, calibration, source commit, and evaluation records were not available in this repository, so this card does not certify those details. The repository name is retained as supplied; robot_type is null in the original metadata.

Payload bytes are preserved: frame data, task mapping, episode metadata, statistics, and videos are not rewritten. Both main and v3.0 carry the complete dataset and this card; existing revision history and tags are retained. Structural checks verify episode counts, frame counts, and data/video path references. This repair does not claim a fresh video-decode, simulator replay, model evaluation, or hardware test.

License

No explicit reusable dataset license was supplied. license: other records that status and does not grant new rights. Existing source and third-party rights still apply.