Explore
For your most demanding E2E robotics workloads.
Explore the platform
Use casesHumanoid robotics
Basketball task
WorkloadsSimulation · Data
Policy training
Evaluation
SoftwareScene & robot assets
Simulation environments
Orchestration
InfrastructureHGX + RTX
Shared storage
Networking
Our fleet is balanced across HGX and RTX architectures. A shared foundation from scene construction through evaluation.
Build scenes in the cloud with the responsiveness of local hardware. RTX PRO 6000 Blackwell GPUs give you a fully ray-traced (RT) viewport you can manipulate in real time — not the degraded preview you get from traditional AI infra lacking RT cores. Your simulation stack, assets, and saved scenes persist between sessions, so you can spin down compute to save costs and pick up where you left off.
02 / 04
Replicate your scene across multiple RTX PRO 6000 nodes based on the scale of your data generation needs, operate your own randomization logic, and release the compute capacity when the run finishes. Every node writes to the same NVMe namespace, so what comes out of generation is already sitting where training reads it — no consolidation step, no copying terabytes between systems.
03 / 04
Pre-train world models and vision-language-action (VLA) models, then adapt robot policies through reinforcement learning (RL). Choose Blackwell and Rubin HGX configurations for your model and training scale. Bring together RTX PRO 6000 simulation, policy inference, and training capacity for RL post-training, with resources sized to each workload and released when the run finishes.
04 / 04
Run GPU-accelerated policy inference alongside RTX-powered simulation, and scale to thousands of trials as needed. Use the same checkpoints, assets, and results throughout offline evaluation, simulation, and validation on connected robot hardware.
Scroll to follow the workflow
Keep the workload moving, from the first simulation to a build ready for the robot.
NVIDIA RTX PRO 6000 Blackwell with 96 GB of memory — enough to hold production-scale scenes.
Scenes and assets load up to 10x faster than a cold pull from object storage — your assets persist on local NVMe between sessions.
Isaac Sim, Isaac Lab, and your preferred stack arrive preconfigured. No driver hunting, no CUDA version mismatches.
Flexibly scale up to tens of thousands of nodes. And spin them down just as easily. Pay for the render hours, not the underlying hardware.
Shared NVMe storage across all nodes means that generated data lands as a training-ready dataset, not multiple partial ones to merge.
Scenes move from interactive to headless without conversion — same USD (Universal Scene Description), same stack, same storage. Point your randomization script at it and scale out.
Match GPU memory and interconnect bandwidth to your training needs. Scale from focused fine-tuning jobs to distributed foundation-model pre-training with NVLink-connected GPUs and high-speed networking.
Scale each part of the RL workflow as needed: RTX for simulation, HGX for policy execution and training
Shared NVMe storage provides high IOPS and sustained throughput for parallel data loading and checkpoint writes. Training reads directly from your generated datasets, while checkpoints remain accessible across simulation, inference, and training.
Scale closed-loop policy evaluation. Run thousands of simulated trials on RTX PRO 6000 Blackwell GPUs, with high-memory compute for policy inference.
Support hardware-in-the-loop testing. Connect deployment hardware (e.g. NVIDIA Jetson) to cloud simulation to measure inference latency and control performance.
Support sim-to-real validation. Compare simulated and real robot performance with shared access to checkpoints, rollout videos, telemetry, and evaluation
Enable robotics foundation model and world model training. Compute, data, and validation matched to each application.
Train whole-body and manipulation policies from demonstrations and simulated experience. Evaluate checkpoints across tasks and robot configurations.
Generate sensor data for assembly and machine tending. Train manipulation policies and test precision across fixtures, parts, and contact conditions.
Train navigation and picking policies across warehouse layouts. Evaluate perception and multi-robot coordination under changing traffic and inventory.
Train perception and planning models on recorded and synthetic sensor data. Evaluate rare scenarios and measure inference performance on target hardware.
Train outdoor navigation and manipulation policies. Test against variations in terrain, lighting, vegetation, and weather.
Train perception models for defect detection and policies for autonomous inspection. Evaluate sensor coverage, navigation, and inference constraints.
Explore the infrastructure behind these workloads.
Explore the platform