In this paper, we present a novel approach for safe autonomous robot navigation in pedestrian-dense environments. We integrate uncertainty estimation into robot actions based on deep reinforcement learning (DRL), considering alleatoric, epistemic, and prediction uncertainties. We integrate observation-dependent variance (ODV) and dropout into the PPO algorithm, and compare and analyze deep ensembles with Monte Carlo dropout (MC-dropout) to estimate policy uncertainty. In uncertain decision-making situations, we shift the robot's social behavior to conservative collision avoidance. Experimental results demonstrate improved learning performance of PPO with ODV and dropout, and the impact of learning scenarios on generalization. MC-dropout is shown to be more sensitive to perturbations and better correlates uncertainty types with perturbations. Through safe action selection, the robot can navigate with fewer collisions in perturbed environments.
