Planning Approximate Exploration Trajectories for Model-Free Reinforcement Learning in Contact-Rich Manipulation

Recent progress in deep reinforcement learning has enabled simulated agents to learn complex behavior policies from scratch, but their data complexity often prohibits real-world applications. The learning process can be sped up by expert demonstrations but those can be costly to acquire. We demonstrate that it is possible to employ model-free deep reinforcement learning combined with planning to quickly generate informative data for a manipulation task. In particular, we use an approximate trajectory optimization approach for global exploration based on an upper confidence bound of the advantage function. The advantage is approximated by a network for Q-learning with separately updated streams for state value and advantage that allows ensembles to approximate model uncertainty for one stream only. We evaluate our method on new extensions to the classical peg-in-hole task, one of which is only solvable by active usage of contacts between peg tips and holes. The experimental evaluation suggests that our method explores more relevant areas of the environment and finds exemplar solutions faster-both on a real robot and in simulation. Combining our exploration with learning from demonstration outperforms state-of-the-art model-free reinforcement learning in terms of convergence speed for contact-rich manipulation tasks.