ROVE: Unlocking Human Interventions for Humanoid Manipulation via Reinforcement Learning
Researchers have introduced ROVE, a reinforcement learning framework designed to improve humanoid manipulation by effectively utilizing imperfect human interventions. The system addresses challenges in collecting high-quality intervention data by employing Optimistic Value Estimation (OVE) to prioritize valuable actions from mixed-quality trajectories. ROVE also incorporates cross-embodiment human experience videos to enhance supervision for failure and recovery modes, ultimately outperforming existing baselines on complex manipulation tasks. AI
IMPACT Enhances humanoid robot capabilities by improving learning from human feedback, potentially accelerating real-world applications.