We'll be at GTC Berlin, October 20–22

NVIDIA GTC Berlin runs from October 20 to 22, and part of our team will be there for all three days. We are not presenting in Berlin. Our last poster, on windshield damage detection, was at GTC 2026 in San Jose. This time we will be in the sessions and taking meetings. If you work on training data for robots, vehicles, inspection or medical imaging and will be in Berlin that week, we would like to meet.

Tuesday the 20th is for full-day training workshops. Jensen Huang gives the keynote at the Tempodrom on Wednesday the 21st, from 11:00 to 13:00 CEST, and the sessions and exhibit hall run at STATION-Berlin on Wednesday afternoon and Thursday. NVIDIA lists more than 100 sessions, with physical AI and robotics, industrial AI and open models and datasets among the main topics. It has not said what the keynote will announce. In-person passes are sold out and only a waitlist is open, but the keynote livestream is free and needs no registration.

The sessions on our list

We sorted the program by the stage of the synthetic data pipeline each session touches. These are the sessions we plan to cover between us, with NVIDIA's session codes for the catalog; a few overlap, so we will split them across the team. For the wider view, Rev Lebaredian, NVIDIA's VP of Physical AI Simulation, gives the talk Building Europe's Autonomous Future With Open Models and Physical AI on Wednesday at 15:30 (SB61077).

Building the scene

In our pipeline, much of the work happens before rendering starts: getting CAD geometry, materials and sensors into a scene that simulates correctly. Tuesday's workshop Build Agent-First SimReady Pipelines for Smart Factory Digital Twins (BW61067) works through that for a factory, including connecting defect detections, synthetic data and human-reviewed agent recommendations to an operational digital twin. In-person seats for the workshops are sold out. On Thursday at 10:00, the Omniverse and Metropolis teams take questions on factory-scale digital twins and industrial AI agents (CWEB1060).

Generating variation

Thursday's Cosmos 3 training lab (BT61197, 13:00 to 14:45, at the Aletto Hotel) is the session closest to our daily work. It uses Cosmos 3 Nano Reasoner to write augmentation prompts from real robot video, generates physically plausible video to scale training data without new recording sessions, and runs a video critic that checks whether the generated video obeys physical laws. The critic is the part we want to see in detail. We wrote about Cosmos 3 when it shipped in June.

NVIDIA's robotics team runs a Connect With the Experts session on agent skills (CWEB1059, Wednesday at 14:30 and Thursday at 14:00). It is an open Q&A on running synthetic data generation, simulation, policy training and evaluation in Isaac Sim and Isaac Lab as repeatable tasks that agents can execute, and on where those agent skills still have gaps. The Thursday 09:00 panel From Generating Pixels to Understanding Worlds (SB61219), with Black Forest Labs, LTX and Vesoma, asks what the open world model ecosystem should standardize first: data, model interfaces, representations, post-training tools or evaluation.

Checking the result

The safety and certification sessions are where training data has to come with evidence. Wednesday's panel Europe's Robotaxi Moment (SB61157, 16:30), moderated by NVIDIA's Ali Kani with TÜV Rheinland, Mercedes-Benz, Wayve and Uber, covers where Europe's robotaxi programs stand between validated prototype and commercial L4 deployment, and how a common open platform can lower certification cost. On Thursday at 14:00, Riccardo Mariani, NVIDIA's VP of Industry Safety, presents NVIDIA Halos (SB61074), including the Halos AI Systems Inspection Lab and how teams build traceable evidence for certification readiness. That is close to what the EU AI Act's Article 10 asks of training data: where it came from, what it covers and how it was checked.

Meeting us

We are in Berlin from Tuesday to Thursday, mostly in the sessions above or around STATION-Berlin, with time between sessions for meetings. Write to hi@datadoo.ai with a line on what you are working on and which day suits you, and we will set a time. It helps to know whether your question is about synthetic data that did not transfer to real conditions, or about the training-data evidence a certifier will ask for.