NVIDIA has emerged as the Autonomous Grand Challenge winner at this week´s Computer Vision and Pattern Recognition (CVPR) conference in Nashville, Tennessee, marking its second straight victory in the End-to-End Driving at Scale category. The achievement, announced at the Embodied Intelligence for Autonomous Systems on the Horizon Workshop, represents the third consecutive year that NVIDIA has taken home an Autonomous Grand Challenge award at CVPR, reflecting a strong track record in advancing self-driving vehicle technologies.
The 2025 challenge, themed ´towards generalizable embodied systems,´ utilized the NAVSIM v2 simulation framework to push boundaries in the development of robust autonomous vehicles. Researchers faced tasks requiring the generation of safe and adaptive driving trajectories within semi-reactive simulations, where dynamic traffic creates unpredictable, real-world-like variables. Evaluation hinged on the Extended Predictive Driver Model Score, which assesses safety, comfort, compliance, and the ability to generalize across synthetic and real scenarios. NVIDIA´s key to winning was its Generalized Trajectory Scoring (GTRS) methodology—a novel approach that produces a wide variety of potential trajectories and uses a transformer decoder, distilled from perception-focused metrics, to filter and select the most effective option for each scenario.
GTRS blends broad, coarse trajectory coverage for a range of driving cases with fine-grained, safety-critical responses, leveraging diffusion policy models conditioned on environmental context. This method supports highly adaptive and robust autonomous driving, outperforming previous benchmarks and demonstrating strong generalization to complex conditions. More broadly, NVIDIA´s presence at CVPR 2025 extends beyond this competition. With over 60 accepted research papers, NVIDIA is advancing innovation in areas spanning automotive, healthcare, robotics, and beyond. The company´s research at CVPR includes breakthroughs in stereo depth estimation, monocular motion understanding, 3D scene reconstruction, closed-loop planning, vision-language modeling, and generative simulation—all essential for the next generation of safer and more adaptable autonomous vehicles. NVIDIA employees are also prominent in workshops and tutorials dedicated to foundation models, simulation, and multi-agent embodied systems, solidifying the company’s influence in the Artificial Intelligence for autonomous systems research community.
