Breaking

NVIDIA Pushes Physical AI Training Onto GTC Berlin 2026 Agenda

September 22, 2026 at 18:21 EDT

Foundation Models
AI Agents
Robotics

NVIDIA is steering its GTC Berlin 2026 training program toward physical AI, highlighting hands-on labs that teach developers to build visual AI agents, work with NVIDIA Cosmos 3, and post-train humanoid robot policies. The push turns GTC Berlin from a conventional conference agenda into a practical curriculum for developers building systems that must understand, simulate and act in the physical world.

Sep. 22, 2026 · NVIDIA

Physical AI moves into the lab

GTC Berlin is being framed around practical training for Cosmos visual agents, Cosmos 3 world models and GR00T humanoid policies, with labs that connect video understanding, simulation and robot deployment.

The strongest scale signal: Cosmos 3 model size

Cosmos 3 Nano

16B

Cosmos 3 Super

64B

Super is four times the parameter count of Nano, matching the split between workstation inference and datacenter deployment.

two-hour labs across Oct. 21-22

100+

sessions across the AI stack

€357

starting pass price

100x

faster long-video summaries claimed for VSS

Hands-on path

Video
Search, summarize and alert from natural language

→

World model
Reason about robot video and create synthetic data

→

Robot policy
Post-train, evaluate and deploy on Jetson Thor

Why builders care

Open checkpoints, training scripts and deployment recipes reduce the distance from prototype video agents to simulation-tested robot policies.

What still needs caution

Physically complex outputs can show artifacts, and real-world deployments still require domain data, GPU capacity and safety validation.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.