ADAS, AV, and AI Meetup - September 17, 2026

Join our virtual meetup to hear talks from experts on cutting-edge topics across AI, ML, and computer vision.

Physical AI systems are increasingly constrained not by model architectures, but by the availability of scalable, high-fidelity real-world data. This talk explores how Dynamic Map Platform transforms survey-grade road assets collected across 1.8 million km of roads worldwide into training- and simulation-ready datasets, including point clouds, imagery, HD maps, road surface models, and 3D Gaussian Splatting representations.

We will discuss why geometric accuracy, semantic understanding, and real-world diversity are critical to building robust autonomous driving systems. Attendees will learn how real-world geospatial data can be structured and scaled for AI training and simulation workflows.

ADAS and autonomous vehicle systems rely on increasingly complex data from cameras, video, LiDAR, radar, and other sensor streams. In this session, Murilo will introduce Voxel51 and explore how the latest multimodal capabilities in FiftyOne help teams bring these data sources together to better understand their datasets and model behavior. He’ll discuss how unified workflows for visualization, search, curation, and evaluation can help ADAS and AV teams uncover challenging scenarios, investigate model failures, and build safer, more reliable autonomous systems.

This is a dataset question rather than a model question. On SLAB's SPEED+ benchmark - synthetic renders for training, real photographs of a satellite mockup under two lighting rigs for test - I built a photometric augmentation stack in good faith to anticipate harsh orbital lighting; it improved the diffusely lit domain on every seed (1.17 → 0.99 rad rotation error) and moved the harshly lit one no further than seed-to-seed variation.

The explanation was visible in the data before any training run: the total variation between the pooled pixel-intensity distribution of my augmented training images and each real domain was 0.63 for the domain that improved and 0.80 for the one that didn't, so the augmentation had pushed training toward the domain it was already least wrong about. I'll also cover where that statistic fails - a pooled marginal can't separate a smooth bright blob from a hard specular edge, which is exactly the spatial distinction the mechanism turns on, so gradient statistics or embedding-space comparison would test it properly.

The question to ask of a robustness intervention is not how much it helps, but which way it pushes your training distribution, and that is answerable before you spend the GPU hours.