Skip to main content

The Physical AI System Architecture

Introduction​

A Physical AI system integrates multiple components into a cohesive architecture that enables intelligent physical behavior.

Learning Objectives:

  • Understand the canonical robotics architecture
  • Explore the sense-plan-act loop
  • Identify integration challenges

Theory​

Sense-Plan-Act Architecture​

┌─────────┐     ┌──────────┐     ┌─────────┐
│ SENSING │ --> │ PLANNING │ --> │ ACTING  │
└─────────┘     └──────────┘     └─────────┘
     ^                                  |
     └──────────── Feedback ────────────┘

Layer Breakdown​

1. Sensing Layer

  • Cameras (RGB, depth, stereo)
  • LiDAR and range sensors
  • IMU (Inertial Measurement Unit)
  • Force/torque sensors
  • Tactile sensors
  • Audio microphones

2. Perception Layer

  • Object detection & tracking
  • Semantic segmentation
  • SLAM (Simultaneous Localization and Mapping)
  • Pose estimation
  • Scene understanding

3. Planning Layer

  • Path planning
  • Motion planning
  • Task planning
  • Grasp planning
  • High-level reasoning (LLM integration)

4. Control Layer

  • Motor controllers
  • Balance controllers
  • Trajectory following
  • Compliance control

5. Actuation Layer

  • Motors (servos, stepper, brushless)
  • Grippers and hands
  • Wheels/legs
  • Tool interfaces

System Integration​

ROS 2 as Middleware​

  • Node-based architecture
  • Message passing
  • Service calls
  • Action servers
  • Parameter management

Real-Time Considerations​

  • Control loops: 100-1000 Hz
  • Perception: 10-60 Hz
  • Planning: 1-10 Hz
  • High-level reasoning: 0.1-1 Hz

Exercises​

  1. Draw a sense-plan-act diagram for a robot vacuum
  2. What happens if sensing fails? Planning fails?
  3. Why do different subsystems run at different frequencies?

Summary​

Physical AI architectures integrate sensing, perception, planning, control, and actuation into real-time systems that enable robots to operate in the physical world.

Further Reading​