Physical AI Has a Data Problem — and It Isn’t Collection Workshop

This online workshop shows Physical AI teams how to turn raw, multi-sensor robot recordings into a curated, searchable corpus using MCAP and FiftyOne. ## From recordings to reliable datasets Physical AI still depends on familiar computer vision tasks such as detection, segmentation, depth estimation and tracking. The challenge is that the basic unit of data has changed. Instead of one image and one label, teams now work with long episodes containing cameras, LiDAR, GPS, IMU and logs,all operating on independent clocks and without simple frame boundaries. Taking place on Wednesday 9 September 2026, from 9:00–10:00 AM PST, this practical online session examines why conventional computer vision tooling often struggles with that complexity. It is organised by the London AI, Machine Learning and Computer Vision Meetup and hosted by super organiser Jimmy G. The Meetup listing describes it as a network event with 116 attendees participating across 48 hosting groups. ## What the workshop covers The session follows the data journey from raw recording to curated corpus. Participants will: - Learn what MCAP is, how it is structured and why it has become ROS 2’s default log format since Iron. - Tour real Physical AI datasets produced by driving, aquatic and forest robots. - Open a complete episode with every sensor synchronised on a shared timeline,including channels that existing software may not know how to decode. - See how FiftyOne 1.19 can open MCAP files natively alongside images and video. - Explore an entire collection of recordings rather than inspecting files one at a time. - Examine smoothness, sensor-health and outlier metrics, including what each signal measures and where it can be misleading. - Learn how to turn quality scores into defensible decisions about what data to keep, investigate or exclude.… Register for this event