Notes on Physical AI &
synthetic data
Notes on Physical AI and synthetic data from the datadoo research team.
Page 1 of 3

Notes from GTC Berlin 2026
We are back from GTC Berlin, and NVIDIA has started on the half of the pipeline it used to leave to everyone else: checking the result. Data Factory blueprints with glass defects, simulation checkpoints, Halos SAFE, and what they leave to the team that ships the model.

We'll be at GTC Berlin, October 20–22
Part of our team will be at NVIDIA GTC in Berlin from October 20 to 22, attending sessions and taking meetings. The sessions on our list, sorted by stage of the synthetic data pipeline, and how to set up a meeting.

Notes on Skild S1 and NVIDIA's September data releases
Skild AI's S1 learns an unseen task from one video, roughly 380 teleoperated demonstrations saved by Skild's count. Notes on that release, NVIDIA's NuRec and robotaxi posts from the same weeks, and the quality-control ratio Skild published alongside.

The open Physical AI stack, read from the data side
NVIDIA now ships an open model for each layer of Physical AI: Cosmos 3 generates worlds, GR00T acts in them, Nemotron 3 runs the agents around them. A plain map of the stack as it stands in August 2026, and the data work all three layers still assume you bring.

What SIGGRAPH 2026 means for synthetic training data
NVIDIA's SIGGRAPH 2026 program, read as a synthetic data release: a 4B world model with open weights running on workstation GPUs, and papers that automate scene reconstruction, materials, and motion. Generating data keeps getting cheaper. Knowing whether it transfers is still the expensive part.

Cosmos 3 thinks before it renders
At GTC Taipei, NVIDIA shipped Cosmos 3: a world model that reasons about a scene before generating it. It folds the generate-and-render core of a synthetic-data pipeline into one model. The validation half, which decides whether the data trains a production model or quietly breaks it, did not move.