ABC Box

Built for open-source behavior cloning. Compatible with ABC out of the box.

ABC Box is a bimanual manipulator designed to work seamlessly with ABC, the fully open-source stack for behavior cloning from UC Berkeley, MIT, Amazon FAR, and XDOF. Whether you're training on the 3,500-hour ABC-130K dataset, fine-tuning ABC-DiT or ABC-VLA checkpoints, or running your own teleoperation collection, [Product Name] drops into the ABC pipeline without custom integration work.

Matched kinematics. ABC Box's two arms mirror the follower configuration used to collect ABC-130K, so policies trained on the released dataset and checkpoints transfer to your hardware with minimal calibration.

GELLO-compatible teleoperation. Pair ABC Box with passive, encoder-only leader arms or YAM leader arms — the same approach ABC uses for both initial demonstrations and DAgger-style intervention collection. 

Sim-to-real aligned. Joint limits, workspace geometry, and camera mount points match the MuJoCo simulation models released alongside ABC, so you can iterate in ABC Sim before deploying to physical hardware.

Three-camera ready. Wrist and top-down camera mounts are pre-configured for ABC's standard 3-camera vision input, matching the observation format expected by both ABC-DiT and ABC-VLA.

Built-in safety and onboard compute. A hardware e-stop gives operators an immediate physical cutoff during teleoperation and evaluation, while onboard compute handles data logging and synchronization during collection — no separate host machine required just to record demonstrations.

Specifications 

What's in the box

• Two follower arms with mounting hardware
• Camera mount kit (wrist × 2, overhead × 1)
• 3 x RSD405 Cameras
• Control interface cable and quick-start guide
• Configuration files pre-tuned for the ABC training and inference pipeline

ABC is an independent open-source research project (Berkeley, MIT, Amazon FAR, XDOF); [Product Name] is a third-party hardware platform designed for compatibility with it.