![[article image]](https://cacm.acm.org/system/assets/0003/3853/012919_Science_Magazine_robotraining.large.jpg?1548784055&1548784055 "A simulated robot learning to walk and avoid obstacles.")
Uber AI Labs researchers have designed an algorithm that organizes its own curriculum for teaching simulated robots to walk on uneven ground without falling.
The Paired Open-Ended Trailblazer (POET) artificial intelligence (AI) initially develops a set of unique terrains, each inhabited by a computer-controlled character that must learn to walk using only two legs and a laserlike rangefinder.
After some practice, the AI revises the obstacle course's level of difficulty, and occasionally inserts a different walker to see whether skills are transferable.
The researchers think POET could enable walkers to eventually traverse rough terrain that they could not have learned without the earlier courses.
They also found POET outperformed an AI that simply raised the difficulty of terrain over time, without trying many indirect paths; the implication is that POET could help real-life robots solve complex tasks, or enable autonomous cars to handle emergencies that programmers overlook.
From Science
View Full Article
Abstracts Copyright © 2019 SmithBucklin, Washington, DC, USA
