microsoft/BusyBox

microsoft/BusyBox

Language: Jupyter Notebook

License: MIT

Stars: 22

Forks: 1

Open issues: 0

Created: 2025-09-05T06:30:53Z

Pushed: 2026-06-21T01:05:05Z

Default branch: main

Fork: no

Archived: no

README:

BusyBox: Benchmarking Affordance Generalization

BusyBox is a physical 3D-printable device for benchmarking affordance generalization in robot foundation models.


It features

• Modular design with 6 interchangeable modules (buttons, switches, sliders, wires, knob, and display)
• Open-source CAD files and bill of materials for easy reproduction
• Optional electronics and Raspberry Pi instrumentation for automated state logging
• Reconfigurable setups enabling systematic evaluation of generalization
• A language-annotated dataset of 1000+ demonstration trajectories oof BusyBox affordances

Please check out our website for more details.

BusyBox assembly instructions

For fully building a instrumented BusyBox capable of state logging, see the [BOM](BOM.md).

First print the BusyBox following [Printing Instructions](cad/printing_instructions.md) with details on files to print and any details on print settings.

Electronic Assembly:

TODO: add instructions on how to assemble electronics with pictures

Firmware Flashing:

Instructions for flashing the Arduino Nano's firmware: [Flashing Firmware](devices/flashing_firmware.md)

Data Collection

See [Data Collection](assets/taskbox_data_collection.docx) for details on our data collection methodology.

Recording Episodes with Aloha

The robots/aloha/ directory contains a complete data recording pipeline for collecting teleoperated demonstrations using the Aloha bimanual robot.

Prerequisites

• Aloha robot hardware (leader + follower arms) set up and calibrated
• USB foot pedal for episode control
• (Optional) Instrumented BusyBox with MQTT bridge for state logging

Quick Start

pip install -r requirements.txt
python robots/aloha/record_busybox_episodes.py

Foot Pedal Controls

| Pedal | Action | |-------|--------| | Left | Start recording | | Middle | Stop and save episode / Refresh task prompt | | Right | Pause/resume teleop / Reject recording |

Configuration

Edit robots/aloha/utils/config.py to set:

• dataset_dir — where episodes are saved
• camera_names — which cameras to record
• using_instrumented_busybox — enable MQTT state logging

Episode Data Format (HDF5)

Each episode is saved as an HDF5 file with:

• /observations/qpos — joint positions (14,) per timestep
• /observations/qvel — joint velocities (14,)
• /observations/effort — joint torques (14,)
• /observations/images/{cam_name} — JPEG-compressed camera frames
• /action — leader arm joint commands (14,)
• /observations/timestamp, /action_timestamp — timing data

Utility Scripts

| Script | Purpose | |--------|---------| | scripts/count_episodes_collected_today.py | Count episodes recorded in the current session | | scripts/busybox_calibration.py | BusyBox sensor calibration | | scripts/visualize_hdf5.ipynb | Visualize recorded episode data | | robots/aloha/eval_rollouts.py | Evaluate policy rollouts |

Raspberry Pi MQTT Bridge

For instrumented BusyBox setups, the devices/pi_sw/ directory contains the MQTT bridge that runs on the Raspberry Pi:

# On the Raspberry Pi
pip install pyserial paho-mqtt
python devices/pi_sw/mqtt_bridge.py --broker-host

See [Flashing Firmware: Installing MQTT on the Pi](devices/flashing_firmware.md#6-installing-mqtt-on-the-pi) for setup details.