Compositional Transfer in Hierarchical Reinforcement Learning
This addresses data-efficiency problems for robotics applications using hierarchical reinforcement learning, representing an incremental improvement with novel mechanisms for compositional transfer.
The paper tackles the high data requirements of reinforcement learning in robotics by introducing Regularized Hierarchical Policy Optimization (RHPO), which uses compositional inductive biases to share data across tasks and controllers. The method demonstrated substantial data-efficiency and performance gains in a week-long physical robot stacking experiment.
The successful application of general reinforcement learning algorithms to real-world robotics applications is often limited by their high data requirements. We introduce Regularized Hierarchical Policy Optimization (RHPO) to improve data-efficiency for domains with multiple dominant tasks and ultimately reduce required platform time. To this end, we employ compositional inductive biases on multiple levels and corresponding mechanisms for sharing off-policy transition data across low-level controllers and tasks as well as scheduling of tasks. The presented algorithm enables stable and fast learning for complex, real-world domains in the parallel multitask and sequential transfer case. We show that the investigated types of hierarchy enable positive transfer while partially mitigating negative interference and evaluate the benefits of additional incentives for efficient, compositional task solutions in single task domains. Finally, we demonstrate substantial data-efficiency and final performance gains over competitive baselines in a week-long, physical robot stacking experiment.