GraphWorld: Long-Horizon Planning with World Models for End-to-End Autonomous Driving

A new autonomous-driving framework called GraphWorld aims to overcome a key limitation of current systems: their inability to model long-term temporal dependencies, which restricts generalization and safety in complex traffic scenarios [1]. The framework, detailed in a paper submitted to arXiv on June 15, 2026, proposes an end-to-end architecture that enhances long-horizon planning through latent world modeling [1]. Most existing end-to-end autonomous driving methods unify perception, prediction, and planning but remain confined to short-horizon decision-making [1]. GraphWorld addresses this by introducing an Ego-Centric Interaction Graph that adaptively models critical neighboring agents based on spatial proximity and propagates relational context to planning queries via cross-node cross-attention [1]. A second component, World-State-Conditioned Planning, learns ego-centric latent world representations by modeling interactions between the ego vehicle and surrounding agents, capturing interaction dynamics and safety-relevant semantics to guide trajectory planning [1]. The researchers evaluated GraphWorld on the Bench2Drive, NAVSIMv1/2, and nuScenes datasets, reporting significant reductions in collision rates and improved long-horizon planning performance [1]. The work arrives as the autonomous-driving industry scales real-world deployments. Waymo, a subsidiary of Alphabet, operates public robotaxi services in 10 U.S. metropolitan areas with 3,871 vehicles providing 500,000 paid rides per week as of June 2026 [6]. The company logged 200 million fully autonomous miles, though it also faces federal investigations into incidents involving stopped school buses and a collision with a child in a school zone [6]. Related research on world models for driving simulation has tackled complementary challenges. A separate framework, VectorWorld, was proposed to enable stable, real-time closed-loop rollouts exceeding one kilometer by generating ego-centric vector-graph tiles incrementally and using a physics-aligned non-ego policy to reduce compounding kinematic errors [5]. GraphWorld’s authors argue that conditioning planning on a latent world state allows interaction-consistent reasoning over extended horizons, rather than decoding trajectories solely from planning queries [3]. The paper is hosted on arXiv under the Computer Vision and Pattern Recognition category [1].