The high probability of hardware failures prevents many advanced robots (e.g. legged robots) to be confidently deployed in real-world situations (e.g post-disaster rescue). Instead of attempting to diagnose the failure(s), robots could adapt by trial-and-error in order to be able to complete their tasks. However, the best trial-and-error algorithms for robotics are all episodic: between each trial, the robot needs to be put back in the same state, that is, the robot is not learning autonomously. In this paper, we introduce a novel learning algorithm called "Reset-free Trial-and-Error" (RTE) that allows robots to recover from damage while completing their tasks. We evaluate it on a hexapod robot that is damaged in several ways (e.g. a missing leg, a shortened leg, etc.) and whose objective is to reach a sequence of targets in an arena. Our experiments show that the robot can recover most of its locomotion abilities in a few minutes, in an environment with obstacles, and without any human intervention. Overall, this new algorithm makes it possible to contemplate sending robots to places that are truly too dangerous for humans and in which robots cannot be rescued.
from cs.AI updates on arXiv.org http://ift.tt/2ebYpTs
via IFTTT
No comments:
Post a Comment