OmniTacTune: Policy-Agnostic Real-World RL for Tactile Residual Adaptation of Visual Policies

OmniTacTune presents a policy-agnostic real-world reinforcement learning pipeline that adds tactile feedback to pretrained visual policies through a lightweight residual policy. The method first bootstraps tactile learning from autonomous rollouts of the base visual policy, then improves the residual through online interaction. According to the paper summary, across four real-world contact-rich tasks, it raises success rates from 5–40% to 85–100% within 40–80 minutes. The reported experiments also cover multiple visual base policies and tactile representations, indicating a practical route for adapting scalable visual priors to manipulation tasks where contact geometry and local force matter.