Simulate fast
Reduce expensive, risky physical testing with realistic simulations, compressing weeks of testing into days.
Generate virtual worlds to reduce hardware testing
Use DeepMind’s Genie 3 or NVIDIA Cosmos—‘world models’—to give Physical Agents an expansive digital sandbox to try, fail and improve. Input text, image, video, and movement to guide the output, which can then be uploaded into simulated environments to make them more realistic.
Simulate the effects of gravity, motion, and energy
MuJoCo-Warp is an open-source engine for robotics R&D. Developed by Google DeepMind and NVIDIA, it offers up to 100x faster simulation (versus standard MuJoCo) to efficiently validate your hardware's physics. MuJoCo-Warp can be used on its own, or paired with NVIDIA Isaac Sim as a core rendering platform.
Try it yourself: Read the MuJoCo documentation or take a tutorial.
Test risky or expensive scenarios
Test millions of edge cases by creating high-fidelity, multi-model synthetic media using Google’s foundation models such as Veo, Nano Banana, Chirp and Lyria models or 200+ other open and third-party models available in Model Garden and deployable on Google Kubernetes Engine (GKE).
Try it yourself: Take a guided tutorial to create synthetic data in Gemini Enterprise Agent Platform, or read NVIDIA's documentation on how to integrate synthetic data into NVIDIA Isaac Sim.
Convert synthetic media into training-ready data
Use our low-code interface and pre-trained models for object detection, depth sensing and segmentation on Gemini Enterprise Agent Platform to convert synthetic media into labeled data in minutes, without manual tagging.
Access workload-optimized GPUs
Run GPU-based simulations on G4 VMs (NVIDIA RTX PRO 6000 Blackwell) to get 9x higher price-performance than our previous generation. It offers up to 768GB of GPU memory and fourth-generation Ray Tracing cores, ideal for the high-performance requirements of simulation.
Learn how: WPP accelerates humanoid robot training 10x with G4 VMs
Simulate fast
Reduce expensive, risky physical testing with realistic simulations, compressing weeks of testing into days.
Generate virtual worlds to reduce hardware testing
Use DeepMind’s Genie 3 or NVIDIA Cosmos—‘world models’—to give Physical Agents an expansive digital sandbox to try, fail and improve. Input text, image, video, and movement to guide the output, which can then be uploaded into simulated environments to make them more realistic.
Simulate the effects of gravity, motion, and energy
MuJoCo-Warp is an open-source engine for robotics R&D. Developed by Google DeepMind and NVIDIA, it offers up to 100x faster simulation (versus standard MuJoCo) to efficiently validate your hardware's physics. MuJoCo-Warp can be used on its own, or paired with NVIDIA Isaac Sim as a core rendering platform.
Try it yourself: Read the MuJoCo documentation or take a tutorial.
Test risky or expensive scenarios
Test millions of edge cases by creating high-fidelity, multi-model synthetic media using Google’s foundation models such as Veo, Nano Banana, Chirp and Lyria models or 200+ other open and third-party models available in Model Garden and deployable on Google Kubernetes Engine (GKE).
Try it yourself: Take a guided tutorial to create synthetic data in Gemini Enterprise Agent Platform, or read NVIDIA's documentation on how to integrate synthetic data into NVIDIA Isaac Sim.
Convert synthetic media into training-ready data
Use our low-code interface and pre-trained models for object detection, depth sensing and segmentation on Gemini Enterprise Agent Platform to convert synthetic media into labeled data in minutes, without manual tagging.
Access workload-optimized GPUs
Run GPU-based simulations on G4 VMs (NVIDIA RTX PRO 6000 Blackwell) to get 9x higher price-performance than our previous generation. It offers up to 768GB of GPU memory and fourth-generation Ray Tracing cores, ideal for the high-performance requirements of simulation.
Learn how: WPP accelerates humanoid robot training 10x with G4 VMs
Train smart
Improve model accuracy by creating reinforcement learning loops enriched with real-world and AI-generated data at a lower cost.
Simplify your entire ML workflow
Simplify your entire ML workflow with Agent Platform, use AutoML to quickly prototype models and explore new datasets without coding expertise, or run massive, Slurm-based distributed training jobs with Cluster Director.
Safely ingest AI-generated code
Use LLMs and agentic tools to automate and scale reinforcement learning while mitigating vulnerabilities like data loss, exfiltration or damage to production systems. GKE Agent Sandbox ingests untrusted code in an isolated environment while maintaining <1s spin-up latency.
Try it yourself: Read the GKE Agent Sandbox documentation or blog.
Continuously detect and predict anomalies
Simplify management and maintenance for training clusters
Get the same MLOps software powering Google’s own training clusters. Get a topological view of cluster health and utilization, always-on health checks, straggler detection, and Elastic Training to automatically reset, swap and scale down degraded nodes without interrupting training runs on Gemini Enterprise Agent Platform, Cluster Director, or GKE.
Reduce the baseline cost of compute
Flexible consumption options like committed use discounts, Dynamic Workload Scheduler, and Spot VMs can deliver the capacity you need while cutting costs by up to 91%.
Learn more: Choose a consumption option
Access a wide range of AI-optimized accelerators with minimal switching costs
Optimize for granular, workload-level objectives, running simulations on NVIDIA Omniverse with G4 VMs, then either NVIDIA GPUs or Cloud TPUs for training and inference, with minimal code rewrites for PyTorch users thanks to native TPU integration.
Train smart
Improve model accuracy by creating reinforcement learning loops enriched with real-world and AI-generated data at a lower cost.
Simplify your entire ML workflow
Simplify your entire ML workflow with Agent Platform, use AutoML to quickly prototype models and explore new datasets without coding expertise, or run massive, Slurm-based distributed training jobs with Cluster Director.
Safely ingest AI-generated code
Use LLMs and agentic tools to automate and scale reinforcement learning while mitigating vulnerabilities like data loss, exfiltration or damage to production systems. GKE Agent Sandbox ingests untrusted code in an isolated environment while maintaining <1s spin-up latency.
Try it yourself: Read the GKE Agent Sandbox documentation or blog.
Continuously detect and predict anomalies
Simplify management and maintenance for training clusters
Get the same MLOps software powering Google’s own training clusters. Get a topological view of cluster health and utilization, always-on health checks, straggler detection, and Elastic Training to automatically reset, swap and scale down degraded nodes without interrupting training runs on Gemini Enterprise Agent Platform, Cluster Director, or GKE.
Reduce the baseline cost of compute
Flexible consumption options like committed use discounts, Dynamic Workload Scheduler, and Spot VMs can deliver the capacity you need while cutting costs by up to 91%.
Learn more: Choose a consumption option
Access a wide range of AI-optimized accelerators with minimal switching costs
Optimize for granular, workload-level objectives, running simulations on NVIDIA Omniverse with G4 VMs, then either NVIDIA GPUs or Cloud TPUs for training and inference, with minimal code rewrites for PyTorch users thanks to native TPU integration.
Operate easily
Use pre-built models from Google DeepMind, or run your own models and agents with automated cluster maintenance and management fit for exabyte scale.
Skip training with pre-built models from Google DeepMind
Give your physical agents the ability to perceive, reason and act without building a model from scratch using Gemini Robotics-ER 1.6 in Google AI Studio, combining DeepMind's breakthroughs in general purpose Robotics with NVIDIA GPUs and Google TPUs on AI Hypercomputer.
Give your robot a voice
Connect edge devices to Gemini Live to process millions of interactions while detecting mood, context, and nuance across 60+ languages.
Manage thousands of agents across hybrid and multicloud environments
In a hybrid, multicloud world, our open platform gives you the freedom to innovate without getting locked in. Google is the creator of Kubernetes—the industry standard for containers—and today Google Kubernetes Engine (GKE) can scale to over 130,000 nodes. It natively integrates with Gemini Enterprise Agent Platform in the cloud and Google Distributed Cloud at the edge, establishing a single containerized pipeline across every environment and cloud provider with virtually limitless scale.
Connect distributed agents in minutes, not weeks
Cross-Cloud Network can help you establish high-bandwidth dedicated connectivity between Google Cloud and other cloud service providers. It's trusted by over 65% of the Fortune 100 to move over 27 exabytes of data per month.
Make on-device functions safer and faster
When you operate devices at the edge, they need to react to their surroundings quickly and safely.
Deploy applications on local infrastructure, managed from the cloud with Google Distributed Cloud, reducing cloud-to-edge round-trip latency from 200ms to sub-10ms and offloading computationally heavy tasks like 3D mapping (SLAM), path planning, and object recognition to on-site hardware. You can even run real-time inference and control loops directly on the robot with Coral Edge NPUs.
Try it yourself: Read the Google Distributed Cloud documentation, including a step-by-step guide for installation.
Operate easily
Use pre-built models from Google DeepMind, or run your own models and agents with automated cluster maintenance and management fit for exabyte scale.
Skip training with pre-built models from Google DeepMind
Give your physical agents the ability to perceive, reason and act without building a model from scratch using Gemini Robotics-ER 1.6 in Google AI Studio, combining DeepMind's breakthroughs in general purpose Robotics with NVIDIA GPUs and Google TPUs on AI Hypercomputer.
Give your robot a voice
Connect edge devices to Gemini Live to process millions of interactions while detecting mood, context, and nuance across 60+ languages.
Manage thousands of agents across hybrid and multicloud environments
In a hybrid, multicloud world, our open platform gives you the freedom to innovate without getting locked in. Google is the creator of Kubernetes—the industry standard for containers—and today Google Kubernetes Engine (GKE) can scale to over 130,000 nodes. It natively integrates with Gemini Enterprise Agent Platform in the cloud and Google Distributed Cloud at the edge, establishing a single containerized pipeline across every environment and cloud provider with virtually limitless scale.
Connect distributed agents in minutes, not weeks
Cross-Cloud Network can help you establish high-bandwidth dedicated connectivity between Google Cloud and other cloud service providers. It's trusted by over 65% of the Fortune 100 to move over 27 exabytes of data per month.
Make on-device functions safer and faster
When you operate devices at the edge, they need to react to their surroundings quickly and safely.
Deploy applications on local infrastructure, managed from the cloud with Google Distributed Cloud, reducing cloud-to-edge round-trip latency from 200ms to sub-10ms and offloading computationally heavy tasks like 3D mapping (SLAM), path planning, and object recognition to on-site hardware. You can even run real-time inference and control loops directly on the robot with Coral Edge NPUs.
Try it yourself: Read the Google Distributed Cloud documentation, including a step-by-step guide for installation.
See how our customers are innovating with physical agents
See how our customers are innovating with physical agents
