Abstract
We demonstrate the first reinforcement-learning application for robotsequipped with an event camera. Because of the considerably lower latency of theevent camera, it is possible to achieve much faster control of robots comparedwith the existing vision-based reinforcement-learning applications usingstandard cameras. To handle a stream of events for reinforcement learning, weintroduced an image-like feature and demonstrated the feasibility of trainingan agent in a simulator for two tasks: fast collision avoidance and obstacletracking. Finally, we set up a robot with an event camera in the real world andthen transferred the agent trained in the simulator, resulting in successfulfast avoidance of randomly thrown objects. Incorporating event camera intoreinforcement learning opens new possibilities for various roboticsapplications that require swift control, such as autonomous vehicles anddrones, through end-to-end learning approaches.