Ant Twowalk demo from Frans, Ho, and Schulman (2018) "Meta Learning Shared Hierarchies"
Full paper: [ Ссылка ]
More on Kevin Frans and OpenAI:
[ Ссылка ]
Frans’s new paper, with Schulman and three others affiliated with the University of California Berkeley, reports new progress on this problem. “If it could get solved it could be a really big deal for robotics but also other elements of AI,” Frans says. He developed an algorithm that helped virtual legged robots learn which limb movements could be applied to multiple tasks, such as walking and crawling. In tests, it helped virtual robots with two and four legs adapt to new tasks, including navigating mazes, more quickly. A video released by OpenAI shows an ant-like robot in those tests. The work has been submitted to ICLR, one of the top conferences in machine learning. "Kevin's paper provides a fresh approach to the problem, and some results that go beyond anything demonstrated previously," Schulman says
![](https://s2.save4k.ru/pic/9dXz93EzWJI/mqdefault.jpg)