NVIDIA Clinches Back-to-Back Victory in CVPR Autonomous Driving Grand Challenge
NVIDIA has once again emerged victorious in the prestigious Autonomous Grand Challenge at the 2025 Computer Vision and Pattern Recognition (CVPR) conference, held this week in Nashville, Tennessee. The recognition was announced during the “Embodied Intelligence for Autonomous Systems on the Horizon” workshop.
This marks the second consecutive year NVIDIA has secured the top spot in the End-to-End Driving at Scale category, and the third year in a row it has claimed a Grand Challenge award at CVPR — a testament to its consistent leadership in autonomous driving research.
This year’s challenge focused on the theme “Towards Generalizable Embodied Systems” and was powered by NAVSIM v2, a sophisticated, data-driven, nonreactive simulation platform for autonomous vehicles (AVs). The competition encouraged participants to develop models capable of handling unexpected and diverse scenarios — going beyond training on human driving data alone.
Developing Safe and Adaptive AV Trajectories
The challenge required teams to generate driving trajectories using multi-sensor input in a semi-reactive simulation. In this setup, the ego vehicle’s route was predetermined, while surrounding traffic moved dynamically — demanding robust and adaptive planning.
Submissions were assessed using the Extended Predictive Driver Model Score, which evaluates performance across four key dimensions: safety, comfort, compliance with traffic rules, and generalization to both real-world and synthetic environments.
NVIDIA’s winning solution demonstrated exceptional capability in generating safe, human-like driving behavior under unpredictable conditions — advancing the goal of smarter, safer autonomous vehicles.