LiDAR Dreamer: Efficient World Model for Autonomous Racing with Cartesian-Polar Encoding and Lightweight State-Space Cells

Download PDF
Open Webpage

Myeongjun Kim, Jong-Chan Park, Sang-Min Choi, Gun-Woo Kim

Published: 14 Oct 2025, Last Modified: 23 Jan 2026InformationEveryoneRevisionsCC BY-SA 4.0
Abstract: Autonomous racing serves as a challenging testbed that exposes the limitations of perception-decision-control algorithms in extreme high-speed environments, revealing safety gaps not addressed in existing autonomous driving research. However, traditional control techniques (e.g., FGM and MPC) and reinforcement learning-based approaches (including model-free and Dreamer variants) struggle to simultaneously satisfy sample efficiency, prediction reliability, and real-time control performance, making them difficult to apply in actual high-speed racing environments. To address these challenges, we propose LiDAR Dreamer, a novel world model specialized for LiDAR sensor data. LiDAR Dreamer introduces three core techniques: (1) efficient point cloud preprocessing and encoding via Cartesian Polar Bar Charts, (2) Light Structured State-Space Cells (LS3C) that reduce RSSM parameters by 14.2% while preserving key dynamic information, and (3) a Displacement Covariance Distance divergence function, which enhances both learning stability and expressiveness. Experiments in PyBullet F1TENTH simulation environments demonstrate that LiDAR Dreamer achieves competitive performance across different track complexities. On the Austria track with complex corners, it reaches 90% of DreamerV3’s performance (1.14 vs. 1.27 progress) while using 81.7% fewer parameters. On the simpler Columbia track, while model-free methods achieve higher absolute performance, LiDAR Dreamer shows improved sample efficiency compared to baseline Dreamer models, converging faster to stable performance. The Treitlstrasse environment results demonstrate comparable performance to baseline methods. Furthermore, beyond the 14.2% RSSM parameter reduction, reward loss converged more stably without spikes, improving overall training efficiency and stability.