Life-Long Learning to Drive by Semi-Supervised Reinforcement Learning


At a glance

Some organizations have reached significant milestones close to or already in on-road deployment. Even so, some real-world challenges linger and cast doubts about the on-road performance of self-driving vehicles. It is not an easy task for an autonomous agent to learn to deal with unencountered and tricky situations that are not foreseen in the design phase. This proposed study aims at tackling this challenge along with its life-long driving. In this study, we investigate a machine learning approach, semi-supervised reinforcement learning (SSRL), in which a continuous life-long learning process enables the autonomous agent to become adaptive and skillful. The agent is trained with available yet insufficient data where a reward function is provided, and then further trained with an algorithm that resembles inverse reinforcement learning, to simultaneously learn a reward and a more general policy in the wider range of unlabeled settings.

principal investigatorsresearchersthemes
Pin Wang
Ching-Yao Chan
Pieter Abbeel
 Semi-Supervised Reinforcement Learning, Life-Long Learning, Autonomous driving behavior