ALOHA Unleashed: a two-armed robot that ties shoelaces and hangs shirts
In one sentence On 12 September 2024 Google DeepMind presents ALOHA Unleashed: a policy trained with Diffusion Policy on teleoperated ALOHA 2 demonstrations that ties shoelaces, hangs a shirt, replaces another robot's finger, inserts gears and tidies kitchen items.
On 12 September 2024 Google DeepMind presented two pieces of work on robot dexterity. The first is ALOHA Unleashed, for two-armed robots.
The method is imitation learning: a person remotely guides the two arms of the ALOHA 2 robot through a difficult task, and these demonstrations become the training data. The model then predicts the robot's actions by starting from random noise and refining it, the way image generators do (Diffusion Policy).
The tasks shown are tying shoelaces, hanging a shirt on a hanger, replacing a finger on another robot, inserting gears and stacking randomly placed kitchen items. According to the authors the policy handles many different shirt types, including one that was not in the training data.
On the same day DeepMind presented DemoStart, which teaches a multi-fingered robot hand mostly in simulation.
Companies
Google DeepMind
Tools
ALOHA Unleashed, Diffusion Policy, ALOHA 2, DemoStart
Tags
Sources