Skip to main content

Physical AI Infrastructure

Physical AI Infrastructure covers the complete workflow from demonstration capture to model deployment: record real robot operations, organize them into trainable datasets, train a model, and deploy it for inference so the robot can learn and execute tasks from demonstrations.

Supported Platform
  • The Linux x86 version supports the complete Physical AI Infrastructure module.
  • The Windows version supports Data Collection, Dataset Management, and Model Training.

Workflow​

  • Data Collection: manage collection tasks, configure devices, and record real-time operations to generate training data segments.
  • Dataset Management: organize collected data, preview segments, assess quality, run cloud conversion, and publish to Hugging Face.
  • Model Training: select converted data artifacts, configure the model and training specification, and track training results.
  • Inference Deployment: load a trained model, configure cameras and robotic arms, and run online inference tasks.

Usage Tips​

  • Keep camera roles and robotic arm configuration consistent across data collection, model training, and inference deployment to reduce input differences between stages.
  • Before starting data collection, confirm that robot serial ports, camera feeds, and the storage directory are all configured.
  • Before training, preview segments in dataset management and delete records with missing images, interrupted motion, or poor quality.
  • Model training uses cloud resources and account points. For offline or local validation, you can first deploy inference with a downloaded model.