LO-World: Unified BEV Latent Modeling for Joint 4D Occupancy and LiDAR Forecasting
Zhimin Liao, Ping Wei, Kai Mao, Shuaijia Chen
Xi'an Jiaotong University
Published November 9, 2026 · on the feed November 11, 2026 · ▲ 1 · arXiv 2611.11912
The abstract
LiDAR point clouds and 3D semantic occupancy represent two distinct modalities capturing fine-grained metric geometry and dense volumetric semantics, respectively, holding natural potential for mutual complementarity. However, existing world models typically forecast these modalities independently or cascade them heuristically, missing the opportunity for reciprocal cross-modal regularization. To address this, we propose LO-World, a unified 4D world model that forecasts both modalities within a shared BEV latent space. This single-stream formulation enables dense semantics and metric geometry to mutually reinforce each other, while eliminating multi-pipeline redundancy. Specifically, LO-World couples an efficient hybrid range-voxel autoencoder with a flow-matching generator that caches historical context to reduce iterative sampling latency. Extensive experiments demonstrate that LO-World outperforms prior state-of-the-art baselines, improving occupancy forecasting IoU by 10.9% and reducing LiDAR Chamfer Distance by 30.1%. Meanwhile, it sustains a real-time throughput exceeding 25~FPS on a single RTX 4090. Our code is available at https://github.com/lzzzzzm/LO-World.
The Brief
1:56A short podcast summary of the paper. Written and voiced by AI.
Transcript
LO-World. Unified BEV Latent Modeling for Joint 4D Occupancy and LiDAR Forecasting. By Zhimin Liao and three others, from Xi'an Jiaotong University. Published on the ninth of November.
Two modalities describe a 3D scene in different ways. LiDAR point clouds capture fine-grained metric geometry: exactly where surfaces are. 3D semantic occupancy captures dense volumetric semantics: what fills each part of the volume. Each holds what the other lacks, and that gives them a natural potential to complement each other. Existing world models, however, usually forecast the two independently, or cascade one into the other heuristically. Either way they miss the chance for reciprocal cross-modal regularization, where each modality keeps the other honest.
LO-World is a unified 4D world model that forecasts both modalities inside one shared BEV latent space. The formulation is single-stream. Rather than two forecasters running side by side, there is one, and its latent has to account for semantics and geometry at once. That is what lets dense semantics and metric geometry reinforce each other, and it also removes the redundancy of maintaining several pipelines.
Specifically, two components make it up. An efficient hybrid range-voxel autoencoder is coupled with a flow-matching generator. Flow matching samples iteratively, and to keep that from becoming slow, the generator caches historical context, which reduces the latency of each round of sampling.
Extensive experiments compare LO-World with prior state-of-the-art baselines. It improves occupancy forecasting IoU by ten point nine percent, and it reduces LiDAR Chamfer Distance by thirty point one percent. Both modalities get better at once. Meanwhile it sustains a real-time throughput of more than twenty-five frames per second on a single RTX 4090.
The design bet is that two views of the same scene are better forecast together than apart. One latent space carries both, the shared stream lets each modality regularize the other, and the reported numbers improve on both sides, at real-time speed on one graphics card.
More with an account
Listen to every day’s papers
Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.
Join free