Blien2002/ShakeBench-tasks

ShakeBench

Overview

ShakeBench contains seven Panda manipulation datasets, stored as separate LeRobot v2.1 directories. Select a dataset directory to load its metadata and episode files; the repository root is not a single flat LeRobot dataset.

Together, the seven currently published directories contain 2,400 episodes and 830,175 frames. Every directory declares a train split; no validation or test split is published.

Published datasets

Dataset directory	Task and episode balance	Episodes	Frames	FPS	Split	Camera features
shakebench_pickplace_300	Pick up an apple, mug, or food can and place it in the target crate; 100 each	300	58,340	20	train, 0:300	main + wrist, RGB 256 × 256
shakebench_push_t_400	Push the light wooden T onto the textured dark gray T target without lifting; one instruction	400	225,384	20	train, 0:400	main + wrist, RGB 256 × 256
shakebench_ring_on_peg_200	Stack the blue large ring, then the yellow small ring on the wooden peg; one instruction	200	92,380	20	train, 0:200	main + wrist, RGB 256 × 256
shakebench_stack_blocks_400	Stack green on blue, then yellow on green; plain blocks and maximum-tolerance square/round tenons, 200 each	400	193,074	20	train, 0:400	main + wrist, RGB 256 × 256
shakebench_upright_400	Stand a mug, wine bottle, boxed drink, or power drill upright; 100 each	400	81,818	20	train, 0:400	main + wrist, RGB 256 × 256
shakebench_panel_operation_400	Press button, toggle up, toggle down, knob on, or knob off; 80 each	400	74,494	20	train, 0:400	main + wrist, RGB 256 × 256
shakebench_wine_rack_300	Place bottles in left, middle, or right slot; 100 per slot instruction	300	104,685	20	train, 0:300	main + wrist, RGB 256 × 256

The table lists all dataset directories currently published in this repository. The task_index values and task counts in each directory's meta/tasks.jsonl and meta/info.json are directory-local, not global task IDs. Wine Rack has three slot instructions and six episode variants: two bottle types (bottled_water_16 and wine_3) × three rack locations, with 50 episodes per bottle/location pair.

File layout

Each dataset directory contains one Parquet file per episode under data/chunk-000/ and LeRobot metadata under meta/:

<dataset-directory>/
  data/chunk-000/episode_000000.parquet
  ...
  meta/info.json
  meta/episodes.jsonl
  meta/episodes_stats.jsonl
  meta/modality.json
  meta/tasks.jsonl
  meta/shakebench_collection.json   # present in six directories

The Push T directory also contains shakebench_sft_subset.json and shakebench_source_collection.json. All seven meta/info.json files declare total_videos: 0 and video_path: null. Main and wrist RGB image bytes are embedded in each episode Parquet file, so loading these datasets does not require separate video files.

Data dictionary

Feature	Type and shape	Meaning
observation.images.main	image, 256 × 256 × 3	Main camera RGB frame
observation.images.wrist	image, 256 × 256 × 3	Wrist camera RGB frame
observation.state	float32[8]	eef_x, eef_y, eef_z, eef_rx, eef_ry, eef_rz, left_finger_qpos, right_finger_qpos
action	float32[7]	delta_x, delta_y, delta_z, delta_rx, delta_ry, delta_rz, gripper
next.reward	float32[1]	Per-frame reward
next.done	bool[1]	Episode completion flag
next.success	bool[1]	Success flag
timestamp	float32[1]	Episode-relative time in seconds
frame_index	int64[1]	Frame number within an episode
episode_index	int64[1]	Episode number within this dataset directory
index	int64[1]	Frame index in this dataset directory
task_index	int64[1]	Task row in this directory's meta/tasks.jsonl

State and action conventions

The eight-dimensional state is not normalized. Its first three values are end-effector position in the robot_base frame, in metres. The next three are orientation as a rotation vector in radians (not Euler angles). The last two are measured signed gripper slide-joint positions in metres.

The seven-dimensional action is ordered as delta x/y/z, delta rotation x/y/z, and gripper. The first six values are relative OSC pose commands; the final value is an absolute gripper command. Collection configuration specifies the robot_base frame, normalized action bounds of [-1, 1], translation scale 0.05 m, rotation scale 0.5 rad, and gripper convention -1=open, +1=close. The Panel units, scales, and gripper convention were verified against its accepted collection metadata and hash-verified collector configuration.

The main and wrist views correspond to task_close and robot0_eye_in_hand, respectively. The collection configurations specify MuJoCo Warp with the official physics profile and gamma=0.0.

Collection and intended use

Programmatic task oracles collected the simulation trajectories. Attempts that failed task checks or showed physical or motion anomalies were excluded or recollected. The published episodes are successful demonstrations, so the data are selected toward successful behavior and do not represent the distribution of failures. All directories are train data; no held-out evaluation split is included.

Local evaluation code: ShakeBench on GitHub. Refer to that repository's own usage notes for its scope and setup.

Loading

Install LeRobot and the Hugging Face Hub client. The examples below load one selected directory at a time. Change dataset_name to any directory in the table to choose a different task.

Load one episode

This minimal example downloads the metadata and episode 0 only:

from pathlib import Path
from huggingface_hub import snapshot_download
from lerobot.datasets.lerobot_dataset import LeRobotDataset

repo_id = "Blien2002/ShakeBench"
dataset_name = "shakebench_upright_400"

repo_snapshot = Path(snapshot_download(
    repo_id=repo_id,
    repo_type="dataset",
    allow_patterns=[
        f"{dataset_name}/meta/**",
        f"{dataset_name}/data/chunk-000/episode_000000.parquet",
    ],
))

dataset = LeRobotDataset(
    repo_id=repo_id,
    root=repo_snapshot / dataset_name,
    episodes=[0],
    download_videos=False,
)

sample = dataset[0]
print(dataset.meta.info["total_episodes"], sample["observation.images.main"].shape)
print(sample["observation.state"].shape, sample["action"].shape)

Load every episode in one directory

To load the complete selected dataset, download its data/ and meta/ trees and omit the episodes argument. With no episode filter, LeRobot loads every episode in that directory.

from pathlib import Path
from huggingface_hub import snapshot_download
from lerobot.datasets.lerobot_dataset import LeRobotDataset

repo_id = "Blien2002/ShakeBench"
dataset_name = "shakebench_upright_400"  # choose any one directory from the table

repo_snapshot = Path(snapshot_download(
    repo_id=repo_id,
    repo_type="dataset",
    allow_patterns=[
        f"{dataset_name}/data/**",
        f"{dataset_name}/meta/**",
    ],
))

dataset = LeRobotDataset(
    repo_id=repo_id,
    root=repo_snapshot / dataset_name,
    download_videos=False,
)

print("episodes:", dataset.meta.info["total_episodes"])
print("episode metadata rows:", len(dataset.meta.episodes))
print("frames:", len(dataset))

sample = dataset[0]
print(sample["observation.images.main"].shape)
print(sample["observation.state"].shape, sample["action"].shape)

The full-directory loading pattern was validated on the 400-episode Panel export: LeRobot initialized the complete dataset with 400 episodes and 74,494 frames. A minimal episode-0 load was also tested with LeRobot 0.3.3, datasets 3.6.0, and PyArrow 22.0.0.

License and citation

No dataset license has been specified. Public availability does not by itself grant a license for reuse; contact the dataset owner for terms.

Please cite or link to the ShakeBench dataset repository; no separate citation is currently available.