Berlin Physical AI, ML, and Computer Vision Meetup - September 25, 2026

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

Also joining us the next day at MotionLab.Berlin? Register for both nights to save your spot at the September 26 workshop and meetup too.

This talk will explore practical workflows for building, evaluating, and improving modern computer vision systems. We'll dive into real-world approaches to dataset curation, model analysis, multimodal AI workflows, and production-ready vision pipelines using open-source technologies.

The session is designed for engineers, researchers, and AI practitioners looking to better understand how teams are developing and scaling computer vision applications today. Expect practical demos, technical insights, and discussions around the evolving AI tooling ecosystem.

Robotics data arrives in two very different shapes. MCAP files are continuous, lossless logs of everything a robot's sensors recorded, built for playback, SLAM, and sensor fusion. LeRobot datasets are synchronized episodes of camera, state, and action, built for training vision-language-action policies. In this talk, Harpreet takes a tour through real datasets in both formats inside Voxel51, showing what each one contains and what it leaves out.

Along the way he covers how the two formats differ in structure and purpose, what breaks when you load real robot data, and how to inspect episodes, sensor streams, and metadata before training on them. Expect live dataset walkthroughs rather than slides.

Robots are becoming smarter and more affordable—but which automation projects actually deliver a return on investment? Drawing on RemBrain’s real-world experience across delivery, retail, construction, and manufacturing, this presentation reveals why many promising robotics concepts fail to become viable products.

It introduces a practical, skill-based approach to flexible automation and shows how compact AI-powered workstations can achieve payback in as little as 6–12 months. The talk also explores where Vision-Language-Action models create genuine value—and where simpler, proven technologies remain more effective.

Honest, numbers-driven discussion with an algorithm focus

Every new robot, camera, or sensor vendor in physical AI usually means rebuilding the data pipeline from scratch, even though the features you actually need, a calibrated frame, a fused object distance, stay the same. This talk demos mloda, an open-source Python framework where a feature pipeline is written once, and only the reader plugin, the small piece that knows how to read one specific device's raw format, ever changes.

Live, using two small synthetic datasets shaped like two different devices' raw output (no hardware involved), I run the same pipeline against both, swapping only the reader plugin, and show the calibration, fusion, and feature extraction steps running unmodified on either. I also show OpenTelemetry lineage tracing, so when a feature looks wrong, you can trace it straight back to the raw reading that produced it, regardless of which device it came from.

The point: stop rebuilding your data pipeline for every new device, reuse it.