Technology

The stack behind Physical AI

Every Physical AI system runs on three layers: perception, reasoning, and action. Understanding the stack tells you why some systems work and why others fail — and what the real engineering constraints are.

Architecture

Three layers every Physical AI system needs

01

Perception

Cameras, LiDAR, radar, microphones, IMUs, force sensors, GPS. The sensors that let a physical AI system read the world. Without perception, there is no data to reason about.

RGB camerasDepth sensors (LiDAR/ToF)IMU (motion)MicrophonesForce/torque sensorsGPS / UWB positioning
02

Reasoning

AI chips and foundation models that process sensor data and decide what to do. This is where on-device AI vs cloud AI matters most — latency, privacy, and reliability all hinge on where reasoning happens.

Edge AI chips (Jetson, Qualcomm)Vision-language modelsVision-language-action (VLA) modelsDiffusion transformer policiesLLM planning layer
03

Action

Motors, servos, grippers, speakers, displays — the actuators that translate decisions into physical effects. For wearables: haptics, audio output, display updates. For robots: joint torques, gripper forces.

Brushless DC motorsHydraulic actuators (humanoids)GrippersSpeakers / earbudsHaptic feedbackOLED/LCD displays

benned

Technology choices behind Kin

Kin runs on-device. Not because it is a technical constraint, but because a personal AI that knows your habits should not send those habits to a data center. Kin runs on the hardware you already own — your phone, your devices — with your knowledge staying local.

The technology pages in this section document the industry landscape. They also explain the constraints benned is building within.

How the stack connects

From sensor to action — the full signal path

In a deployed Physical AI system, all three layers run simultaneously and in sequence. A camera (perception) feeds frames to an edge AI chip running a vision-language-action model (reasoning), which outputs a motor command executed by a joint actuator (action). In a humanoid robot running GR00T N1 on Jetson Thor, this loop runs at 200 times per second.

For wearables, the loop is: microphone or health sensor (perception) feeds an NPU running a context model (reasoning), which surfaces a notification, adjusts audio, or triggers an alert (action). The constraint changes — not 200Hz motor control, but sub-second response on a 3.5W power budget.

Understanding which layer is the bottleneck in a given deployment is how you diagnose why a Physical AI system underperforms — and what to fix.

Last updated: July 2026