A Pipeline for Uncertainty Estimation in the Visual Perception of Reinforcement Learning Agents

Safety-critical applications using deep neural networks (DNNs) must operate reliably under varying conditions. To ensure this, uncertainty estimation in DNNs becomes essential for enabling safer actions and more robust decision making.

This work introduces a pipeline that allows researchers to focus more on developing uncertainty-aware DNNs by drastically reducing the amount of repetitive setup tasks. The pipeline features a modular design, high flexibility, and uses the Godot game engine for environment simulation. A key innovation is the development of a partial pipeline for the fully automatic generation of annotated image data for Object Detection. Additionally, through a UDP connection, Godot exchanges data with Python, which provides a wrapped Gymnasium environment in Python that one can use to train Reinforcement Learning agents efficiently. The proposed pipeline applies to any custom 3D environment and supports various algorithms and models.