This is a GaTech course research project that aims to investigate different safe reinforcement learning (RL) approaches for a robot to navigate safely in dynamic cluster environments. Here’s the project paper link: Interactive Learning on Safe Navigation in Cluster Dynamic Environments.

Abstract

Safe navigation is essential for mobile autonomous systems to deploy in real-world environments. In this paper, we want to investigate different safe reinforcement learning (RL) approaches for a robot to navigate safely in dynamic cluster environments. Based on the same learning framework where safe action is used, if the RL agent provides an unsafe action, we developed two different approaches: one uses an optimization-based safety controller to produce safe actions, the other uses human interventions as safe actions. Our experiment results indicate that the optimization-based safety controller can safeguard the robot from collision, but the approach using human interventions achieves very similar performance as regular RL.

Methodology

On the left, the proposed reinforcement learning approach with an SSA-based safety controller (safe RL-SC). On the right, the interactive learning approach using human interventions (safe RL-HI).

Results

The mean performance of 30 runs of each approach after 200 episodes of training (Success rate/Collision rate/Halting rate).