Training a deep policy gradient-based neural network with asynchronous learners on a simulated robotic problem

Recent advances in deep reinforcement learning methods have attracted a lot of attention, because of their ability to use raw signals such as video streams as inputs, instead of pre-processed state variables. However, the most popular methods (value-based methods, e.g. deep Q-networks) focus on discrete action spaces (e.g. the left/right buttons), while realistic robotic applications usually require a continuous action space (for example the joint space). Policy gradient methods, such as stochastic policy gradient or deep deterministic policy gradient, propose to overcome this problem by allowing continuous action spaces. Despite their promises, they suffer from long training times as they need huge numbers of interactions to converge. In this paper, we investigate in how far a recent asynchronously parallel actor-critic approach, initially proposed to speed up discrete RL algorithms, could be used for the continuous control of robotic arms. We demonstrate the capabilities of this end-to-end learning algorithm on a simulated 2 degrees-of-freedom robotic arm and discuss its applications to more realistic scenarios.