Comparing High Entropy Reinforcement Learning to Navmesh for Automated Collision Bug Testing

Aurélien Chambon, Nicolas Grelier

Keywords: QA Agents, Reinforcement Learning, Collision bugs, Entropy

TL;DR: A RL-based method for automatic collision bug detection that explores the entire map and notifies the development team.

Abstract: Automatic bug testing in video games is a topic of growing interest for industry and academia. Reinforcement learning (RL) is emerging as a promising approach, particularly for detecting collision bugs, which are tedious and time-consuming for Quality Assurance (QA) analysts to evaluate manually. However, prior works often rely on visual inspection of results of RL navigation agents. In this short paper, we introduce an early-stage method for automating collision bug detection by comparing the traversal time of an RL navigation agent with that of a navmesh agent. We pretrain this agent in an obstacle free environment, deploy it along a route and exploit entropy-driven RL exploration to bypass obstacles, before resetting to the pretrained policy to continue map coverage. This approach enables scalable map analysis for detecting clipping and collision issues, while automatically flagging anomalies for developer review. Additionally, we provide early insights on the impact of high entropy tuning in our method.

Submission Number: 8