A collaboration team from Google Research proposes “PI-ARS: Accelerating Evolution-Learned Visual-Locomotion with Predictive Information Representations”, a new learning algorithm that enables robots to successfully perform difficult visual-locomotion tasks in complex environments.
ARS belongs to the family of Evolution Strategy (ES) algorithms which are based on the concept of natural selection.
There have been numerous works that demonstrate the effectiveness of applying ES algorithms to continuous control problems. However, the scalability of ES to large capacity models is a significant issue.
To tackle this problem, PI-ARS use predictive information (PI) to create a compressed representation of the environment and then applies the Augmented Random Search (ARS) algorithm to transform the learned compact representation into robot actions.
(a) uneven stepping stones, (b) quincuncial piles
The research team showed that the learnt policies could successfully transfer to a real quadruped robot.
PI-ARS is able to learn vision-based policies that successfully solve challenging visual-locomotion tasks like moving over a series of four step stones, separated by gaps.
During the tests it achieved a 100% success rate in a stepping stone environment, much better as previous results of 40% success.
Learn more:
- Research paper: “PI-ARS: Accelerating Evolution-Learned Visual-Locomotion with Predictive Information Representations” (on arXiv)
- Video (on GitHub)
- Story source (on Google Research)
