by Bob Yirka , Tech Xplore
2024년 9월 23일
Google DeepMind unveils two new AI-based robot hand systems—ALOHA Unleashed and DemoStart
Experimental setup: simulated (top) and real (bottom) robot environments and tasks. Credit: arXiv (2024). DOI: 10.48550/arxiv.2409.06613
Engineers working on Google's DeepMind project have announced the development of two new AI-based robot systems. One called ALOHA Unleashed was developed to advance the science of bi-arm manipulation. The other, called DemoStart, was developed to advance the capabilities of robot hands that have multiple fingers, joints, or sensors.
Details for ALOHA Unleashed have been posted on the DeepMind site and also on GitHub. Details for DemoStart have been posted on the arXiv preprint server.
As the research team notes, most robot hands developed to pick up and move objects generally act alone—they have no second hand to help them. In this new effort, the research team used AI technology to teach a robot to use both of its hands in conjunction to complete a "difficult" task, such as tying a shoe. The result is ALOHA Unleashed.
As the team also notes, the new system builds on ALOHA 2 and the ALOHA platform, which was developed at Stanford University for use in tele-operating applications. The new system improves dexterity and also allows two robot hands to become "aware" of one another as they work together on a common problem.
The robot hands were taught via demonstration to do tasks such as hanging a shirt or repairing a robot part. Afterward, diffusion methods were applied to give the robot hands some degree of prediction, helping them anticipate what the other would be doing.
The research team on DemoStart noted that complex dexterity in robots is going to mean using more fingers, joints and sensors than are currently used on most robot hands. To achieve that requires some degree of coordination between them.
Like the ALOHA Unleashed project, coordination required the introduction of AI into the learning process. With DemoStart, they used reinforcement learning to help the robot gain a sense of its abilities when given control of multiple arm, hand and finger joints, in addition to fingertips.
The approach involved giving the robot hands simple tasks and slowly ramping up the difficulty. They found they could teach a two-fingered robot with several joints and sensors to reorient a cube, tighten a nut and neaten a workspace.
More information: Maria Bauza et al, DemoStart: Demonstration-led auto-curriculum applied to sim-to-real with multi-fingered robots, arXiv (2024). DOI: 10.48550/arxiv.2409.06613
ALOHA Unleashed: A Simple Recipe for Robot Dexterity, aloha-unleashed.github.io/asse … /aloha_unleashed.pdf
DeepMind blog: deepmind.google/discover/blog/ … -in-robot-dexterity/
DemoStart: sites.google.com/view/demostart
Journal information: arXiv