Virtual to Real Reinforcement Learning for Autonomous Driving
This addresses the challenge of non-affordable trial-and-error in real environments for autonomous driving, offering a solution for safer and more efficient policy training.
The paper tackles the problem of training autonomous driving policies with reinforcement learning by proposing a novel realistic translation network that converts virtual images to realistic ones, enabling policies trained in simulation to adapt to real-world driving, with experiments showing successful adaptation.
Reinforcement learning is considered as a promising direction for driving policy learning. However, training autonomous driving vehicle with reinforcement learning in real environment involves non-affordable trial-and-error. It is more desirable to first train in a virtual environment and then transfer to the real environment. In this paper, we propose a novel realistic translation network to make model trained in virtual environment be workable in real world. The proposed network can convert non-realistic virtual image input into a realistic one with similar scene structure. Given realistic frames as input, driving policy trained by reinforcement learning can nicely adapt to real world driving. Experiments show that our proposed virtual to real (VR) reinforcement learning (RL) works pretty well. To our knowledge, this is the first successful case of driving policy trained by reinforcement learning that can adapt to real world driving data.