NVIDIA Achieves Back-to-Back Victory in Autonomous Driving Challenge at CVPR, Showcasing Advanced GTRS Method

NVIDIA has once again claimed victory in the End-to-End Driving at Scale category of the Autonomous Grand Challenge at the Computer Vision and Pattern Recognition (CVPR) conference, which took place this week in Nashville, Tennessee. This win marks the company's second consecutive triumph in this category and its third consecutive Autonomous Grand Challenge award at CVPR. The theme of this year’s challenge, "Towards Generalizable Embodied Systems," focused on NAVSIM v2, a data-driven, nonreactive autonomous vehicle (AV) simulation framework. Researchers were tasked with developing systems that could generate driving trajectories from multi-sensor data in a semi-reactive simulation. In such environments, the ego vehicle's plan is set at the start, but background traffic evolves dynamically, simulating real-world conditions where AVs must adapt to unexpected situations. The NVIDIA AV Applied Research Team introduced a novel method called Generalized Trajectory Scoring (GTRS). This method generates a diverse set of potential driving trajectories and systematically filters out the best ones. GTRS combines coarse and fine-grained trajectories, the latter being particularly suited for safety-critical scenarios. These fine-grained trajectories are generated using a diffusion policy conditioned on the environment, while a transformer decoder evaluates them based on safety, comfort, and traffic rule compliance. The decoder captures subtle yet crucial differences between similar trajectories, allowing for robust and adaptive decision-making in a variety of driving conditions. GTRS has demonstrated exceptional performance, achieving state-of-the-art results on challenging benchmarks. This success underscores the method's ability to generalize well across different scenarios, enhancing the reliability and safety of autonomous vehicle systems. The NVIDIA team's approach could significantly accelerate the development of smarter, safer AVs by addressing the complex issue of handling unforeseen events on the road. Beyond this specific challenge, NVIDIA's presence at CVPR 2025 was notable, with over 60 papers accepted across various domains, including automotive, healthcare, and robotics. Three papers—FoundationStereo, Zero-Shot Monocular Scene Flow, and Difix3D+—were even nominated for the Best Paper Award. These papers highlight advancements in stereo depth estimation, monocular motion understanding, 3D reconstruction, closed-loop planning, vision-language modeling, and generative simulation, all of which are essential components in the development of safer, more generalizable AVs. Several workshops and tutorials at CVPR featured prominent NVIDIA researchers, including Marco Pavone, senior director of AV research, and Sanja Fidler, vice president of AI research. These sessions covered topics such as data-driven autonomous driving simulation, open-world 3D scene understanding, safe artificial intelligence for all domains, and foundation models for V2X-based cooperative autonomous driving. Additionally, NVIDIA hosted workshops on full-stack, GPU-based acceleration of deep learning and continuous data cycles using foundation models, further emphasizing its commitment to AI and automotive innovation. Industry experts have praised NVIDIA's contributions to autonomous driving research. Jose Alvarez, director of AV applied research at NVIDIA, noted that GTRS addresses a critical gap in current AV technology by enabling better trajectory prediction and adaptation in dynamic environments. The industry's recognition of NVIDIA's achievements reflects the company's position as a leader in AI and automotive technology. With a global team of hundreds of scientists and engineers, NVIDIA continues to push the boundaries of what is possible in the realm of autonomous vehicles, making significant strides towards a future where self-driving cars are both intelligent and reliable. For those interested in delving deeper, the CVPR conference program includes detailed workshops, tutorials, and papers from NVIDIA, accessible through their official research page. The NVIDIA GTC Paris keynote, delivered by founder and CEO Jensen Huang, also provides valuable insights into the company’s latest advancements in AI and autonomous driving.