The paper proposes a comprehensive framework called MRIC (Model-based Reinforcement-Imitation Learning with Mixture-of-Codebooks) for autonomous driving simulation. The key insights and components are:
Closed-loop differentiable simulation provides meaningful learning signals and achieves efficient credit assignment, but suffers from gradient explosion and weak supervision in low-density regions.
To address these issues, MRIC introduces two policy regularizations:
A dynamic multiplier mechanism is proposed to eliminate interference between the regularizations and the main objective, while ensuring their effectiveness.
A temporally abstracted mixture-of-codebooks module is designed to compress the diverse behaviors of heterogeneous agents into a series of prototype vectors, addressing the issues of prior holes and posterior collapse.
Extensive experiments on the Waymo Open Motion Dataset show that MRIC outperforms state-of-the-art baselines on key metrics like collision rate, minSADE, and time-to-collision JSD, demonstrating its ability to simulate diverse and realistic autonomous driving behaviors.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Baotian He,Y... at arxiv.org 04-30-2024
https://arxiv.org/pdf/2404.18464.pdfDeeper Inquiries