Simulation-Based Reinforcement Learning Policy Optimization for Tactile Manipulation: A Case Study on the Eyesight Hand

Abstract

Robotic manipulation remains a complex and unsolved challenge due to the need for adaptability across diverse objects and tasks. In this work, we explore how to train effective manipulation policies using reinforcement learning in simulation for the Eyesight Hand: a novel, low-cost, tactile-enabled robotic hand. We implement a range of experiments in MuJoCo to evaluate the impact of controller types, observation spaces, reward formulations, and curriculum strategies on policy performance. Our findings highlight the benefits of delta position control, a carefully selected observation space including joint states, control vectors, object pose, and contact forces, and success-driven curriculum learning. Our study establishes baseline strategies for training robust, tactile-based policies on this in-house hardware.