UniIntervene

UniIntervene: Agentic model reduces human interventions in real-world robot RL by 57%

UniIntervene is a proposed agentic intervention model for human-in-the-loop reinforcement learning (HiL-RL) that autonomously detects unproductive exploration and recovers robot policies toward high-value states, replacing the bulk of human corrections. The system uses future-conditioned action-value estimation, a temporal value-risk critic, and a goal-conditioned recovery policy drawing from a memory of past interventions. Experiments on real-world robotic manipulation tasks show a 57% reduction in human interventions and an 8.6% improvement in average success rate over state-of-the-art HiL-RL baselines.