kyounginbaik/anubis_mug_to_sink

Anubis Mug to Sink

This dataset contains LeRobot-format trajectories for the Anubis mobile manipulator performing the Put mug in sink task. It is part of the SimVLA data release and stores synchronized robot state, actions, joint commands, and RGB video observations.

Project page: https://kyounginbaik.github.io/simvla/

Dataset Details

• Format: LeRobot v3.0
• Robot: Anubis
• Task: Put mug in sink
• Split: train
• Episodes: 8,700
• Frames: 9,939,192
• FPS: 20
• Video observations: 320x240 RGB, H.264, no audio
• Modalities: front camera, left wrist camera, right wrist camera, robot state, joint angles, Cartesian/base action, joint action, timestamps, task and episode indices

Repository Structure

data/
  chunk-000/
    file-000.parquet
    ...
meta/
  info.json
  stats.json
  tasks.parquet
  episodes/
videos/
  observation.images.front/
  observation.images.wrist_left/
  observation.images.wrist_right/

The Parquet files contain per-frame state, action, and indexing fields. The image observations are stored as videos under videos/{video_key}/..., following the LeRobot dataset layout.

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("kyounginbaik/anubis_mug_to_sink")
sample = dataset[0]
print(sample.keys())

Feature Summary

Main fields include:

• observation.images.front
• observation.images.wrist_left
• observation.images.wrist_right
• observation.state
• joint_angles
• action
• action.joint
• timestamp
• frame_index
• episode_index
• task_index
• subtask_index
• kitchen metadata fields such as kitchen_num, kitchen_sub_num, kitchen_type, and initial_pose

Intended Use

This dataset is intended for research on robot imitation learning, vision-language-action policy training, synthetic-to-real transfer, and analysis of mobile manipulation behavior in the SimVLA setting.

Citation

If you use this dataset, please cite the SimVLA project.