Physical AI Has a Data Problem. It Isn't Collection - September 9, 2026

Name: Physical AI Has a Data Problem. It Isn't Collection - September 9, 2026
Start: 2026-09-09
End: 2026-09-09

Sep 09, 2026

9:00 AM - 11:00 AM PST

Online. Register for the Zoom!

About this event

Understanding, curating, and debugging multimodal MCAP data in FiftyOne.

Physical AI still relies on familiar computer vision tasks—detection, segmentation, depth, tracking. What's changed is the data unit: no longer a single image and label, but an episode—a dozen sensors ticking on independent clocks for minutes, with no frame boundaries.

Most computer vision tooling assumes the old unit and breaks on the new one.

That's why Physical AI teams end up with buckets of .mcap files nobody can characterize. Recording is cheap, so logs pile up faster than anyone curates them. Ask what's actually in there—which tasks, which conditions, how many failures and of what kind—and the honest answer is usually a shrug.

MCAP has been ROS 2's default log format since Iron, and as of FiftyOne 1.19 it opens natively: cameras, LiDAR, GPS, IMU, and logs on one shared timeline, alongside your images and video.

This workshop goes from raw recording to curated corpus. We'll cover what MCAP is and why it's built that way, tour real Physical AI datasets across driving, aquatic, and forest robots, and open an episode with every sensor synced—including channels nothing knows how to decode.

Then we'll tackle quality, the harder half: what smoothness, sensor-health, and outlier metrics actually measure, where each falls short, and how to turn a score into a defensible decision.

You'll leave knowing how to load your own recordings, query a whole corpus instead of a single file, and which quality signals to trust for which job.

Host

What You'll Walk Away With

How to load your own MCAP recordings into FiftyOne, with every sensor synced on one shared timeline
How to query a whole corpus of episodes instead of inspecting one file at a time
What smoothness, sensor-health, and outlier metrics actually measure, and where each falls short
Which quality signals to trust for which job, and how to turn a score into a defensible decision

Build better physical AI models.

Physical AI Has a Data Problem. It Isn't Collection - September 9, 2026

What You'll Walk Away With