Astral Fellows at CVPR 2025: Poster Session, Industry Booths, and Researchers Who Define the Field

Astral Fellows researchers traveled to Nashville in June 2025 to present their accepted paper at the IEEE/CVF Computer Vision and Pattern Recognition conference, the largest and most competitive computer vision conference in the world.

CVPR 2025 drew over 12,000 researchers to Nashville, Tennessee in June. It is consistently ranked the most cited computer vision conference in the world, and the 2025 edition received more than 13,000 paper submissions, accepting fewer than 25% of them. Walking into the convention center on day one, surrounded by the researchers behind the papers that defined the field over the past decade, was an experience our team will not forget quickly.

The Paper: PhysNav-DG

Our researchers presented PhysNav-DG, a paper on robust VLM-sensor fusion for autonomous navigation. The work introduces a dual-branch architecture that predicts navigation actions from multi-sensor inputs while simultaneously generating chain-of-thought explanations for its decisions. A modified Adaptive Kalman Filter adjusts its noise parameters dynamically based on environmental context, drawing on semantic insights from LLaMA 3.2 11B and BLIP-2 alongside raw sensor streams.

The paper was accepted at the CVPRW 2025 workshop track, part of the full IEEE/CVF Computer Vision and Pattern Recognition conference. The team also introduced MD-NEX, a novel multi-domain benchmark unifying indoor navigation, autonomous driving, and social navigation with ground-truth actions and human-validated explanations. Experiments showed PhysNav-DG improving navigation success rates by over 20% against prior baselines.

The Poster Session

CVPR poster sessions run for hours and function as the real intellectual core of the conference. Researchers cycle through, stop at papers that interest them, ask pointed questions, and compare notes on related work. Our poster drew a sustained crowd. Questions came from PhD students at Berkeley and Carnegie Mellon, researchers from Waymo and Toyota Research Institute, and faculty whose papers we had been reading for years. The Kalman Filter adaptation mechanism and the explainability angle generated the most discussion. Several people asked for the arXiv link before they left.

“I had read maybe a hundred of these researchers' papers before Nashville. Standing at our poster answering their questions about our own work was something I had not let myself imagine was possible.”

Meeting Rene Vidal

One of the genuine highlights of the week was a conversation with Rene Vidal, whose work on subspace clustering, sparse representations, and the mathematical foundations of deep learning has shaped the field for two decades. Vidal is currently a Professor at the University of Pennsylvania and one of the most recognized figures in computer vision research worldwide. We spoke with him briefly about the theoretical underpinnings of our sensor fusion approach and he offered a sharp observation about how the Kalman Filter adaptation relates to certain convergence results in optimization. It was the kind of exchange that only happens at a conference like CVPR, and it reframed how we are thinking about the follow-up work.

Industry Booths

The industry exhibition floor at CVPR is unlike anything you encounter at a smaller conference. Google, Microsoft, Meta, Tencent, NVIDIA, Waymo, Qualcomm, Amazon, Fourier Robotics, and dozens of other organizations set up full booths staffed by their research teams. These are not recruitment tables. They are active research showcases where you can discuss papers, see live demos, and speak directly with the scientists building the systems.

At the Google booth, our team spent time with researchers from Google DeepMind's robotics and perception teams who were demoing real-time 3D scene reconstruction. The conversations connected directly to PhysNav-DG's multi-sensor pipeline and raised interesting questions about how our Kalman Filter adaptation would behave under the kinds of distribution shifts their deployment work encounters.

What CVPR Actually Is

CVPR is not just the largest computer vision conference. By paper citations, it is the most influential academic venue in all of computer science. The 2024 edition received over 11,500 submissions. In 2025 that number grew again. Getting a paper accepted here, even at the workshop level, means the work was reviewed by expert researchers against the same bar that filters out the overwhelming majority of PhD-level submissions globally. Workshop acceptance rates at CVPR typically sit between 25% and 35%.

For high school students to present original research at this conference is genuinely unusual. It is not a category that the conference tracks, because it essentially does not exist as a population. You submit a paper. The reviewers review it. That is the only standard that applies.

Looking Forward

The conversations started in Nashville are continuing. We came back with sharper research questions, several collaborator leads, and a clearer sense of where PhysNav-DG fits in the broader literature on explainable autonomous systems. The follow-up work is already underway, targeting submission windows for upcoming top-tier conferences.