Mitigating Epistemic Uncertainty in Autonomous Vehicle Trajectory Planning via Bayesian Neural Networks

International Journal of Computer Science and Technology | Vol. 5, No. 1, January 2024 | pp. 1–8

DOI: 10.46882/2024/IJCST/000228

Article Type: Original Research Paper

Title: Mitigating Epistemic Uncertainty in Autonomous Vehicle Trajectory Planning via Bayesian Neural Networks

Names of Authors: Klaus Meyer¹, Rachel Green²

Authors’ Affiliations: ¹Autonomous Systems Engineering Center, Fraunhofer Institute, Karlsruhe, Germany; ²Robotics Intelligence Group, Apex Auto Labs, Detroit, USA

Abstract: Autonomous vehicle systems operating within highly dynamic urban settings struggle with sensor noise and unpredictable pedestrian activities. Standard deterministic trajectory planning models fail to quantify model uncertainty, occasionally resulting in unsafe maneuvers under novel traffic scenarios. This paper introduces an adaptive trajectory planning framework that uses Bayesian Neural Networks (BNNs) to estimate epistemic uncertainty fields in real time. The model estimates variance boundaries across its structural network paths, outputting a localized confidence metric alongside each path prediction. If tracking confidence falls below a pre-configured target, the navigation layer selects a defensive path option. We verified this framework on an autonomous driving test platform navigating complex urban simulations. The validation results show that our Bayesian model decreased localized near-miss incidents by 41.2% compared to standard deterministic planning networks. The system executed its processing iterations within an average latency window of 22 ms, satisfying real-time safety requirements.

Keywords: Autonomous Vehicles, Trajectory Planning, Bayesian Neural Networks, Epistemic Uncertainty, Safety-Critical Systems, Robotics

Manuscript Timeline: Received: August 12, 2023; Revised: October 15, 2023; Accepted: November 28, 2023; Published: January 08, 2024

Citation: Meyer, K., & Green, R. (2024). Mitigating Epistemic Uncertainty in Autonomous Vehicle Trajectory Planning via Bayesian Neural Networks. International Journal of Computer Science and Technology, 5(1), 1–8. DOI: 10.46882/2024/IJCST/000228

Subscribe to read the full article: https://internationalscholarsjournals.org/subscribe-to-read