Can Reinforcement Learning for Continuous Control Generalize Across Physics Engines?
This addresses the problem of RL robustness for researchers and practitioners in robotics and simulation, but it is incremental as it builds on existing algorithms without introducing new methods.
The study investigated whether reinforcement learning algorithms for continuous control can generalize across different physics engines, finding that MuJoCo-trained models transfer best while PyBullet-trained ones fail to generalize, and that minimizing random seed effects improves generalizability.
Reinforcement learning (RL) algorithms should learn as much as possible about the environment but not the properties of the physics engines that generate the environment. There are multiple algorithms that solve the task in a physics engine based environment but there is no work done so far to understand if the RL algorithms can generalize across physics engines. In this work, we compare the generalization performance of various deep reinforcement learning algorithms on a variety of control tasks. Our results show that MuJoCo is the best engine to transfer the learning to other engines. On the other hand, none of the algorithms generalize when trained on PyBullet. We also found out that various algorithms have a promising generalizability if the effect of random seeds can be minimized on their performance.