Automated Knowledge Distillation and Iterative Refinement for Active Learning in Robotic Manipulation

**Abstract:** This paper proposes a novel framework for enhancing active learning in robotic manipulation tasks through automated knowledge distillation and iterative refinement. Existing active learning strategies often require substantial human intervention for query selection and reward shaping. We leverage established techniques in knowledge distillation to transfer learned policies from a simulated environment to a real robot, then employ an iterative loop of active exploration, data collection, and policy refinement driven by a learned utility function. This allows the robot to autonomously navigate a manipulation task while simultaneously improving the efficiency of data acquisition and the robustness of the resulting policy. The proposed framework demonstrably outperforms traditional active learning methods in both simulation and real-world experiments, paving the way for more efficient and adaptable robotic systems.