Persistent self-supervised learning principle: Study and demonstration on flying robots
Kevin G. van Hecke · 2015
We introduce, study and demonstrate Persistent Self-Supervised Learning (PSSL), a machine learning method for usage onboard robotic platforms. The PSSL model leverages a standard supervised learning method to simplify the learning problem, but acquires training data in an unsupervised and autonomous manner. Using two platforms, a small multicopter on earth and the space based test bed SPHERES inside the International Space Station , we demonstrate the PSSL principle on a proof of concept problem: learning monocular depth estimation using stereo vision. The robot operates first in a ground truth mode based on the distance perceived by the stereo system, while persistently learning the environment using monocular cues. After the performance of the estimator transcends a ROC quality measure, the robot switches to operation based on the monocular depth estimates. Our results show the viability of the PSSL method, by being able to navigate a room on the basis of learned monocular vision, without collecting any training data beforehand. We identify a major challenge in PSSL caused by a training bias due to behavioral differences in the estimator and the ground truth based operation; however, this is a known problem also for related learning methods such as reinforcement learning. PSSL helps solve this problem by 1) clearly separating the learning problem from the behavior and 2) the possibility to keep learning during estimator behavior.