OmniDrive: An LLM-Choreographed Multi-Agent World Model with Unified Latent Co-Compression for Multi-View Driving Video Generation
Researchers have introduced OmniDrive, a novel LLM-choreographed multi-agent world model designed for generating multi-view driving videos. This system addresses challenges in integrating heterogeneous control inputs and fusing per-camera latent representations by employing a shared symbolic interlingua. The DRIVE-CHOREO framework utilizes three Qwen2.5-VL agents to create a unified, position-aware token sequence that is co-compressed with video data, achieving state-of-the-art results on the nuScenes dataset for multi-view consistency and BEV mAP. AI
IMPACT Introduces a new method for generating realistic driving videos, potentially improving simulation and training for autonomous systems.