ALOHA Unleashed

DeepMind's 2024 work: massive teleoperated data plus a diffusion policy teach a bimanual robot to tie shoelaces.

ALOHA Unleashed is work by Tony Zhao, Ayzaan Wahid, Chelsea Finn, and colleagues at Google DeepMind, published at CoRL 2024, asking how far pure imitation learning alone can push bimanual dexterity. The recipe is deliberately simple — hence the name. On the low-cost bimanual platform ALOHA 2, 35 operators teleoperated under a standardized protocol to collect more than 26,000 demonstrations across 5 real tasks; the policy is a Transformer encoder-decoder with a diffusion loss (that is, a diffusion policy), about 217 million parameters, predicting the next 50 steps of action in one shot. This let the robot autonomously tie shoelaces, hang a shirt on a hanger, and swap a fingertip on another robot. The paper also found that swapping in ACT-style L1 regression, at the same model size, dropped the shirt task's success rate from 70% to 25%.

ExampleIn the shoelace-tying task, the robot first centers the shoe on the table and straightens the laces, then ties a bow. With the shoe centered and the laces already laid flat, success was 70%; with the shoe rotated up to ±45° and the laces unstraightened, success dropped to 40%.

Also called

ALOHA Unleashed: A Simple Recipe for Robot Dexterity

Related

ALOHA 2 · Diffusion Policy · Action Chunking with Transformers · Bimanual Manipulation · Imitation Learning · Mobile ALOHA

Sources

ALOHA Unleashed: A Simple Recipe for Robot Dexterity (arXiv 2410.13126)
ALOHA Unleashed 项目页 (Chinese)
Proceedings of The 8th Conference on Robot Learning, PMLR 270

As of

2024-10