How to train and deploy robots at scale: Google Cloud physical AI

Google Cloud physical AI brings robotics simulation, training and real-world operation together on one platform.

Before a robot can safely pick up a box in a warehouse or operate on a factory floor, engineering teams must generate massive synthetic datasets, train large vision-language-action models and run them on local hardware with split-second latency. Testing directly on physical machines risks broken hardware, downtime and safety incidents. Google Cloud physical AI consolidates simulation, training and edge deployment into one managed pipeline. Softprom is the official partner of Google.

What was announced

Google Cloud unveiled an end-to-end physical AI stack for robotics developers that connects three stages: Simulate, Train and Operate. Teams can run thousands of virtual environments in parallel on G4 VMs powered by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, train VLA and World Model Architecture (WMA) networks on AI Hypercomputer with A3, A4 and A5 machines, and deploy models to the floor using Gemini Robotics ER 2, Gemini Live and Gemma 4.

Why this matters

For CIOs and engineering leaders in manufacturing and logistics, stitching together separate tools for physics simulation, video labeling, GPU training and edge computing is slow and costly. A unified pipeline shortens iteration cycles, reduces hardware risk and feeds real-world sensor logs back into future training runs. The platform also supports hybrid and offline execution via Google Distributed Cloud, so production robots keep working with sub-10 ms local inference even without network connectivity.

Technical details

• Simulate: NVIDIA Isaac Sim, NVIDIA Cosmos and MuJoCo Warp on GKE with G4 VMs.
• Train: AI Hypercomputer with A3, A4, A5 VMs; Cluster Director with managed Slurm; Google Cloud Managed Lustre for sub-millisecond storage latency.
• Data labeling: Gemini plus NVIDIA Cosmos for automatic bounding boxes and VLA annotations.
• Operate: Gemini Robotics ER 2 (preview in Model Garden), Gemini Live with 70+ languages, Gemma 4 for under 10 ms local inference.
• Feedback loop: Cloud Storage, Pub/Sub and BigQuery for petabyte-scale fleet telemetry.

FAQ

What hardware powers the simulation stage?

G4 VMs on Google Kubernetes Engine, accelerated by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, run NVIDIA Isaac Sim, NVIDIA Cosmos and MuJoCo Warp in parallel environments.

How does Google Cloud physical AI handle offline robots?

Google Distributed Cloud runs managed compute on-premises, while Gemma 4 open models execute vision, audio and tool calls locally in under 10 milliseconds without a network connection.

Can existing training workflows reuse this platform?

Yes. Teams can choose Vertex AI managed training, Cluster Director with Slurm, or native GKE orchestration, and store multi-camera datasets on Managed Lustre with sub-millisecond latency.

Softprom and Google

Softprom is the official partner of Google. Our team helps engineering and IT leaders design robotics and AI infrastructure pipelines on Google Cloud.

Prepared by the Softprom editorial team based on vendor materials (Softprom DistriFlow). Source: original article.