Abstract
Achieving safe, reliable real-world robotic manipulation requires agents toevolve beyond vision and incorporate tactile sensing to overcome sensorydeficits and reliance on idealised state information. Despite its potential,the efficacy of tactile sensing in reinforcement learning (RL) remainsinconsistent. We address this by developing self-supervised learning (SSL)methodologies to more effectively harness tactile observations, focusing on ascalable setup of proprioception and sparse binary contacts. We empiricallydemonstrate that sparse binary tactile signals are critical for dexterity,particularly for interactions that proprioceptive control errors do notregister, such as decoupled robot-object motions. Our agents achieve superhumandexterity in complex contact tasks (ball bouncing and Baoding ball rotation).Furthermore, we find that decoupling the SSL memory from the on-policy memorycan improve performance. We release the Robot Tactile Olympiad (RoTO) benchmarkto standardise and promote future research in tactile-based manipulation.Project page: https://elle-miller.github.io/tactile_rl