Project
ALOHA
Low-cost bimanual hardware and teleop stack paired with ACT-style algorithms.
ProjectFeaturedBeginner
ACT (Action Chunking with Transformers)
Action Chunking with Transformers for low-cost bimanual imitation; popular with ALOHA-style hardware.
Why we recommend it
ACT (with ALOHA-style setups) is a practical entry to bimanual imitation on low-cost arms. Choose it when you care about real dual-arm demos more than the newest VLA. Hardware and teleop skill required. Intermediate.
Overview
ACT (Action Chunking with Transformers) popularized transformer policies that predict action chunks for bimanual imitation, especially on low-cost ALOHA hardware. It is one of the fastest paths from teleop demos to a working real-robot policy for tabletop bimanual tasks.
The method is intentionally pragmatic: chunk actions to reduce compounding error, use a CVAE-style latent for multi-modality, and keep the stack simple enough for lab courses and startups. Many later systems (including parts of the LeRobot ecosystem) treat ACT as a must-have baseline.
Who it is for
Teams with ALOHA-like bimanual setups; courses teaching real-robot IL; anyone needing a strong non-diffusion baseline.
Key highlights
- Proven on low-cost bimanual real robots
- Action chunking reduces compounding error
- Simple training recipe widely reimplemented
- Natural pair with ALOHA teleop stacks
When to use
- Bimanual tabletop tasks with teleop demos
- You need a reliable real-robot baseline quickly
- Diffusion inference is too slow for your control loop
When not to use
- Long-horizon mobile manipulation without major extensions
- Language-conditioned generalist VLA pretraining at scale
Getting started
- 1Collect clean bimanual demos at fixed camera extrinsics.
- 2Train ACT with official or LeRobot configs.
- 3Evaluate with frozen lighting and known object sets first.
- 4Only then stress-test distribution shift.
Papers & reading
How it compares
Excellent with ALOHA-style bimanual demos; less general than DP for multi-modal actions.
Caveats & pitfalls
- Sensitive to demo quality and latency alignment.
- Chunk size must match task dynamics.
Content reviewed 2026-07-23