Event

New Framework Combines Transfer Learning and Reward Redesign to Improve DRL Motion Planning

1 reports · 1 sources

Yuanzhi He and other authors propose a task decomposition framework for deep reinforcement learning (DRL) that combines transfer learning (TL) with reward redesign (RR) for training robot manipulation policies in obstacle environments, with the goal of improving learning efficiency and robustness. The framework's transfer strategy selectively transfers only the first-layer critic parameters between structurally related tasks. The authors report that experiments on the PyBullet Panda platform show that the framework achieves consistent improvements in obstacle scenarios; the best-performing policy was deployed on a real Franka Emika FR3 robot as a feasibility demonstration. The deployment on the real robot is only a qualitative demonstration; the report does not provide quantitative results.