Multi-Task Reinforcement Learning with Soft Modularization
This addresses optimization issues in multi-task RL for robotics, though it is incremental as it builds on existing modularization methods.
The paper tackles the challenge of parameter sharing and gradient interference in multi-task reinforcement learning by introducing a soft modularization technique that reconfigures a base policy network for each task, resulting in improved sample efficiency and performance on robotics manipulation tasks.
Multi-task learning is a very challenging problem in reinforcement learning. While training multiple tasks jointly allow the policies to share parameters across different tasks, the optimization problem becomes non-trivial: It remains unclear what parameters in the network should be reused across tasks, and how the gradients from different tasks may interfere with each other. Thus, instead of naively sharing parameters across tasks, we introduce an explicit modularization technique on policy representation to alleviate this optimization issue. Given a base policy network, we design a routing network which estimates different routing strategies to reconfigure the base network for each task. Instead of directly selecting routes for each task, our task-specific policy uses a method called soft modularization to softly combine all the possible routes, which makes it suitable for sequential tasks. We experiment with various robotics manipulation tasks in simulation and show our method improves both sample efficiency and performance over strong baselines by a large margin.