LiSTAR: Ray-Centric World Models for 4D LiDAR Sequences in Autonomous Driving
TLDR
LiSTAR is a generative world model for 4D LiDAR sequences using ray-centric representations and attention for high-fidelity, controllable synthesis.
Reasoning
The paper presents a novel approach with strong quantitative results (e.g., 76% MMD reduction) and addresses key challenges in LiDAR data generation. However, it is limited to LiDAR modality and does not cover interactive or RL-based world models, and the abstract lacks explicit discussion of limitations or broader applicability.
Read-first score
Read-first score 51.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 34.
Field roles
Rank sensitivity
Stability: volatile; rank range: 368.
Keyword Scores
Deep Analysis
Innovations
- Hybrid-Cylindrical-Spherical (HCS) representation to preserve data fidelity by mitigating quantization artifacts common in Cartesian grids
- Spatio-Temporal Attention with Ray-Centric Transformer (START) that explicitly models feature evolution along individual sensor rays for robust temporal coherence
- 4D point cloud-aligned voxel layout for conditioning and discrete Masked Generative START (MaskSTART) framework for efficient, high-resolution, layout-guided compositional generation
Methodology
LiSTAR operates directly on the sensor's native spherical geometry using a Hybrid-Cylindrical-Spherical (HCS) representation to avoid quantization artifacts. It employs a Spatio-Temporal Attention with Ray-Centric Transformer (START) to capture complex dynamics from sparse temporal data by modeling feature evolution along individual sensor rays. For controllable synthesis, it introduces a 4D point cloud-aligned voxel layout and a discrete Masked Generative START (MaskSTART) framework that learns a compact tokenized representation for layout-guided compositional generation.
Key Results
LiSTAR achieves state-of-the-art performance across 4D LiDAR reconstruction, prediction, and conditional generation, with a 76% reduction in generation MMD, a 32% improvement in reconstruction IoU, and a 50% reduction in prediction L1 Med.