The deployment of Reinforcement Learning (RL) on physical robots still stumbles on several challenges, such as sample-efficiency, safety, reproducibility, cost, and software platforms. In this paper, we introduce MoveRL, an environment that exposes a standard OpenAI Gym interface, and allows any off-the-shelf RL agent to control a robot built on ROS, the Robot OS. ROS is the standard abstraction layer used by roboticists, and allows to observe and control both simulated and physical robots. By providing a bridge between the Gym and ROS, our environment allows an easy evaluation of RL algorithms in highly-accurate simulators, or real-world robots, without any change of software. In addition to a Gym-ROS bridge, our environment also leverages MoveIt, a state-of-the-art collision-aware robot motion planner, to prevent the RL agent from executing actions that would lead to a collision. Our experimental results show that a standard PPO agent is able to control a simulated commercial robot arm in an environment with moving obstacles, while almost perfectly avoiding collisions even in the early stages of learning. We also show that the use of MoveIt slightly increases the sample-efficiency of the RL agent. Combined, these results show that RL on robots is possible in a safe way, and that it is possible to leverage state-of-the-art robotic techniques to improve how an RL agent learns. We hope that our environment will allow more (future) RL algorithms to be evaluated on commercial robotic tasks.
|Title of host publication||The 33rd Benelux Conference on Artificial Intelligence and the 30th Belgian Dutch Conference on Machine Learning (BNAIC/BENELEARN 2021)|
|Subtitle of host publication||AI in ACTION Joint International Scientific Conferences on AI|
|Publication status||Published - 2021|
- Safe Reinforcement Learning
- Path Planning