GraphWorld: Long-Horizon Planning with World Models for End-to-End Autonomous Driving
Researchers have developed two new frameworks, Metis and GraphWorld, aimed at improving autonomous driving and urban navigation systems. Metis decouples video generation and action prediction using a Mixture-of-Transformers architecture, enhancing efficiency and generalization. GraphWorld focuses on long-horizon planning by introducing an Ego-Centric Interaction Graph to model agent relationships and guide trajectory planning. Both approaches demonstrate state-of-the-art performance on various benchmarks, reducing collision rates and improving planning capabilities in complex scenarios. AI
IMPACT These models advance long-horizon planning and efficiency in autonomous driving systems, potentially improving safety and generalization in complex scenarios.