Before a robot can safely pick up a box in a warehouse or work on a factory floor, it needs to be trained to do so. This requires engineering teams to generate massive amounts of synthetic data, train large vision and movement models, and run those models on local hardware with split-second response times. Testing directly on physical machines risks broken hardware, costly downtime, and safety hazards.
Today, we are sharing how Google Cloud physical AI brings simulation, model training, and real-world deployment together on one platform for robotics developers and engineering teams.
Instead of stitching together separate tools for physics simulation, video labeling, GPU training, and edge computing, teams can run the entire robotics pipeline on Google Cloud:
Stage
Workload
Google offering
Simulate
NVIDIA Isaac Sim , Mujoco,
Train
Auto labeling video sequences and training VLA / WMA
AI Hypercomputer with A3, A4 & ,A5 machines accelerated by NVIDIA (Cluster Director, GKE, Training clusters on Gemini platform) and Managed Lustre
Operate
Streaming multi-camera, sensor data and using google model
GCS, Pub/Sub, Bigquery, Gemini Live, Gemma 4, GDM ER 2
Each stage connects directly to the next, moving training data from virtual environments into production robots and feeding real-world sensor logs back into future training runs.
1. Simulate: Test robots in virtual environments before touching hardware
To speed up testing new robot behaviors, developers can run thousands of virtual environments in parallel on Google Kubernetes Engine (GKE) using graphics-optimized G4 VMs powered by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. This setup supports physics and simulation tools including NVIDIA Isaac Sim open-simulation framework, NVIDIA Cosmos open world foundation models, and Google DeepMind's accelerated physics engine, MuJoCo Warp.
In these virtual environments, engineering teams can run large-scale reinforcement learning tests, check that multi-camera views stay consistent across angles, and benchmark Vision-Language-Action (VLA) models, which link what a robot sees and hears to how it moves. Teams can more easily simulate their use cases without the downtime, hardware costs, or safety risks of physical trial and error.
2. Train: Auto-label video data and train large robotics models
Training robots typically requires huge amounts of labeled video to learn how the physical world works. Google Cloud pairs the Gemini model family with NVIDIA Cosmos to analyze and label video automatically. Because Gemini processes video, audio, and spatial context across long recordings, the toolchain can automatically create bounding boxes, label objects in a scene, and generate VLA training annotations.
Once the data is labeled, teams can move straight into distributed training on Google Cloud's AI Hypercomputer to build and fine-tune large VLA and World Model Architecture (WMA) networks:
- Fast data storage: Multi-camera video and movement datasets sit on Google Cloud Managed Lustre, a fully managed parallel file system with sub-millisecond latency and terabytes-per-second speed. This keeps GPUs fed with data and speeds up frequent model checkpointing.
Fast data storage: Multi-camera video and movement datasets sit on Google Cloud Managed Lustre, a fully managed parallel file system with sub-millisecond latency and terabytes-per-second speed. This keeps GPUs fed with data and speeds up frequent model checkpointing.
- Direct GPU control: Teams that want bare-metal GPU performance on A3, A4, and A5 VMs can use Cluster Director (for managed Slurm on GCE or GKE) or GKE (natively or via the Slinky Slurm Operator Add-On). It schedules jobs across Google's Jupiter network fabric to keep bandwidth high during large reinforcement learning runs.
Direct GPU control: Teams that want bare-metal GPU performance on A3, A4, and A5 VMs can use Cluster Director (for managed Slurm on GCE or GKE) or GKE (natively or via the Slinky Slurm Operator Add-On). It schedules jobs across Google's Jupiter network fabric to keep bandwidth high during large reinforcement learning runs.
- Managed Training: Teams that prefer not to build and maintain DIY cluster plumbing can use Vertex AI / Training clusters on Gemini platform, which provide managed Slurm orchestration, automated hardware fault recovery, and ready-made recipes for distributed SFT and RL, or serverless custom training jobs for fine tuning.
Managed Training: Teams that prefer not to build and maintain DIY cluster plumbing can use Vertex AI / Training clusters on Gemini platform, which provide managed Slurm orchestration, automated hardware fault recovery, and ready-made recipes for distributed SFT and RL, or serverless custom training jobs for fine tuning.
3. Operate: Run models on the floor and learn from real-world data
Running robots in factories and warehouses requires both cloud-scale planning and fast local execution. Google Cloud helps teams deploy world models and their own custom agents across cloud, hybrid, and offline environments:
- Task planning and voice interaction: Gemini Robotics ER 2 (Embodied Reasoning, now in preview in Model Garden and the Gemini API in Google AI Studio) turns spatial and physical understanding into step-by-step task plans and multi-robot orchestration for lower-level robot controls.
Task planning and voice interaction: Gemini Robotics ER 2 (Embodied Reasoning, now in preview in Model Garden and the Gemini API in Google AI Studio) turns spatial and physical understanding into step-by-step task plans and multi-robot orchestration for lower-level robot controls.
- Alongside it, Gemini Live supports visual intelligence and low-latency voice conversations in over 70 languages, so robots can understand what they see and respond to spoken instructions at the same time.
Alongside it, Gemini Live supports visual intelligence and low-latency voice conversations in over 70 languages, so robots can understand what they see and respond to spoken instructions at the same time.
- On-premises and offline execution: Google Distributed Cloud (GDC) runs managed compute and containers directly on factory and warehouse floors for low-latency camera processing and fleet management. For robots operating without a network connection, lightweight Gemma 4 open models run on local hardware to handle vision, audio, and tool calls in under 10 milliseconds.
On-premises and offline execution: Google Distributed Cloud (GDC) runs managed compute and containers directly on factory and warehouse floors for low-latency camera processing and fleet management. For robots operating without a network connection, lightweight Gemma 4 open models run on local hardware to handle vision, audio, and tool calls in under 10 milliseconds.
- Continuous improvement: Petabytes of fleet logs and sensor streams flow through Cloud Storage and Pub/Sub into BigQuery. Teams can use this data to spot equipment issues early and feed real-world results back into their next round of model training.
Continuous improvement: Petabytes of fleet logs and sensor streams flow through Cloud Storage and Pub/Sub into BigQuery. Teams can use this data to spot equipment issues early and feed real-world results back into their next round of model training.
Get started with Google Cloud Physical AI
Whether you are running your first physics simulation or managing a fleet of warehouse robots, you can start building on Google Cloud today:
- Simulate: Spin up graphics-optimized G4 VMs on Google Kubernetes Engine to run NVIDIA Isaac Sim and MuJoCo-Warp.
Simulate: Spin up graphics-optimized G4 VMs on Google Kubernetes Engine to run NVIDIA Isaac Sim and MuJoCo-Warp.
- Train: Explore AI Hypercomputer training recipes on Gemini platform or set up Cluster Director with Google Cloud Managed Lustre.
Train: Explore AI Hypercomputer training recipes on Gemini platform or set up Cluster Director with Google Cloud Managed Lustre.
- Operate: Try Gemini Robotics ER 2 in preview, or contact your Google Cloud team to set up a Physical AI workshop.
Operate: Try in preview, or contact your Google Cloud team to set up a Physical AI workshop.
- AI infrastructure
- Partners
