Safely navigating complex, interactive situations is one of the major challenges in planning for automated vehicles. The key difficulty is that drivers have unobservable properties determining their driving style. To make optimal decisions, the belief over latent driver states has to be considered, which is computationally challenging. For this reason, online planning algorithms suffer from the curse of dimensionality in highly interactive scenarios. On the other hand, learned policies often have difficulties generalizing to new environments with different driving styles. We propose to use policies trained in belief space as heuristics to guide online belief space planning algorithms, thereby alleviating the curse of dimensionality. We evaluate the proposed approach in a cooperative merging scenario.
Guiding Belief Space Planning with Learned Models for Interactive Merging
2022-10-08
1381690 byte
Conference paper
Electronic Resource
English
Path Planning in Belief Space with Pose SLAM
Springer Verlag | 2017
|Introducing Incremental eXpectation to Belief Space Planning
British Library Conference Proceedings | 2019
|