Skip to main content

System Architecture Summary

Understanding how the different hardware components work together is crucial for building an effective Physical AI system.

Component Roles​

Simulation Workstation Role​

Develops and tests control policies in a simulated environment (Isaac Sim).

The workstation provides:

  • High-fidelity physics simulation
  • Visual rendering of robot environments
  • Safe testing of control algorithms
  • Rapid iteration without hardware risk
  • Data generation for training AI models

Edge Inference Device Role​

Deploys the trained policies from the workstation onto the physical robot for real-world execution.

The edge device handles:

  • Real-time sensor processing
  • AI model inference at low latency
  • Motor control and actuation
  • Local decision-making
  • Safety-critical operations

Data Flow Architecture​

Sensor Data Flow​

The Edge Kit's sensors (camera, IMU) provide real-time data to the Jetson Orin. This data is processed locally for tasks like:

  • SLAM (Simultaneous Localization and Mapping): Building a map of the environment while tracking the robot's position
  • Object Recognition: Identifying objects and obstacles in the environment
  • Depth Perception: Understanding distances and 3D spatial relationships
  • Motion Tracking: Monitoring the robot's orientation and movement

Actuator Command Flow​

The Jetson Orin runs the AI models and sends low-level control commands to the robot's motors and servos.

This includes:

  • Joint Position Commands: Telling motors where to move
  • Velocity Control: Specifying how fast movements should occur
  • Torque Commands: Controlling the force applied by actuators
  • Coordinated Movement: Synchronizing multiple joints for smooth motion

Development Workflow​

  1. Design & Train (Workstation): Develop control policies and train AI models in simulation
  2. Transfer (Network/Storage): Move trained models to the edge device
  3. Deploy (Edge Device): Run the models on the physical robot
  4. Evaluate (Both): Monitor performance and collect data for refinement
  5. Iterate: Use real-world data to improve simulation and retrain models

Why This Architecture Matters​

This two-tier architecture (powerful workstation + efficient edge device) provides the best of both worlds:

  • Safety: Test dangerous or complex behaviors in simulation first
  • Speed: Rapid iteration without waiting for physical hardware
  • Performance: Real-time control with low-latency edge inference
  • Scalability: Train once in simulation, deploy to multiple robots
  • Cost-Effectiveness: Expensive compute only needed for development, not deployment