BusyBox Dataset

Description

The BusyBox dataset is a collection of language-annotated demonstration trajectories for various tasks from the BusyBox benchmark, collected by teleoperating a bimanual mobile Trossen Aloha robot. The dataset is in the LeRobot v2.1 format. It contains several hours of data for the following BusyBox tasks:

• Setting sliders to specified positions
• Pulling out and inserting wires with audiojack-like connectors
• Rotating a knob to a specified position
• Flipping on/off switches
• Reorienting the BusyBox
• Pressing buttons of various colors
• Moving the robot's arms in a prescribed manner

The consolidated version of the dataset contains all of the above.

Sample demonstrations of these tasks are shown in the following video:

BusyBox project page: https://github.com/microsoft/BusyBox

Paper: see the project webpage

Uses

This dataset can be used for finetuning VLAs for the BusyBox tasks as well as in earlier VLA training stages for enhancing the VLA's general capabilities at controlling bimanual robots.

Contacts

Galen Mullins, Andrey Kolobov

Citation

@misc{busybox2025,
  title={Benchmarking Affordance Generalization with BusyBox},
  author={Fortier, Dean and Adamson, Timothy and Hellebrekers, Tess and LaScala, Teresa and Ennin, Kofi and Murray, Michael and Kolobov, Andrey and Mullins, Galen},
  booktitle={Eval&Deploy Workshop at CoRL-2025: Evaluation and Deployment Across the Robot Learning Lifecycle},
  year={2025}
}