Enhancing Tactile-based Reinforcement Learning for Robotic Control

Effectively combining tactile sensing and reinforcement learning (RL) creates powerful new pathways for sophisticated robot manipulation. However, tactile information is not always fully exploited by neural network-based approaches in deep RL due to its unique characteristics (e.g. sparsity). Departing from conventional reliance on idealised state representations, we present a new approach to strengthen the performance of sensory-driven agents for complex manipulation tasks. We provide a novel application and analysis of tailored reconstruction and multi-step dynamics objectives that help the agent more effectively leverage its tactile observations, and propose training these objectives on a separated auxiliary memory. We find that dynamics-based objectives unlock higher-performing agents that are able to predict future contacts with high precision. Experimental results show the efficacy of our approach through a simulated robotic agent on three complex control tasks with touch and proprioception alone.