PolySona: Parameter-Efficient and Modular Latent Behavior Modeling for Traffic Simulation

Laura Yu Zheng, Hamidreza Yaghoubi Araghi, Tony Wu, Tianyi Zhou, Ming Lin

Keywords: latent variable modeling, trajectory prediction, driving style, traffic simulation, lora, vae

TL;DR: We propose a parameter-efficient method for extracting latent behavior representations to modular adapters in a mixture-of-experts-esque framework.

Abstract: In rare but safety-critical driving scenarios, we hypothesize that trajectory outcomes become increasingly multi-modal based on differences between driver style compared to non-critical, common scenarios. However, current approaches for trajectory prediction rarely account for differences in driving style, which may lead to "averaged" driving style in predictions. While average-case behavior may work well in straight driving, easy scenarios, it limits the diversity of outcomes in more complex scenes or in rare events. Extraction of driving style has several benefits, as it enables simulation of counterfactual outcomes in real-world log replays and potentially more accurate predictions through style-consistent predictions. In this paper, we present a parameter-efficient Mixture-of-Experts framework for extraction of latent driving styles in trajectory prediction models. We choose a parameter-efficient approach to reduce forgetting in well-generalized trajectory prediction models, while offering portability of trained driving style modules. We also propose a Style Consistency Metric to quantify how often a model’s multi‐modal outputs cover the true driving style. In our results, we benchmark different mixture-of-LoRA approaches with our method and show qualitative results that show how the learned experts specialize, and how model saliency changes with our approach.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 15087