Physical AI · Layer 2

Everything about Physical AI Hardware

Robots, from desk-mounted arms to humanoids, are built from the same four components: sensors, compute, actuators, and structure. This page goes through each in turn, with verified products and prices. Sensors include RGB cameras, depth, LIDAR, audio, IMU, and force. Compute means the NVIDIA Jetson ladder, from the Orin Nano Super at $249 up to Thor at $1,999+. Actuators include servos, DC motors, steppers, grippers, and rotors. Structure ranges from PLA and PETG to aluminum extrusion and carbon fiber.

Reference   Last updated May 17, 2026

1. Key terms

TermDefinition
SensorA device that collects information from the world. Cameras see, microphones hear, LIDAR measures distances with lasers.
ComputeThe chip / board where AI models run. Jetson, Raspberry Pi, or your Mac.
ActuatorAnything that makes a robot move or act. Servo motors rotate joints, wheels spin, grippers open and close.
Servo motorA motor that rotates to a specific angle and holds. The primary actuator in robotic arms.
TOPSTera Operations Per Second. How fast an AI chip can process. Jetson Orin Nano = 67 TOPS.
CUDANVIDIA's programming platform for GPUs. Nearly all AI code runs on it.
LIDARLight Detection and Ranging. Shoots laser beams and measures bounce-back time to create 3D maps.
IMUInertial Measurement Unit. A small chip that measures tilt, rotation, and acceleration.
CSICamera Serial Interface. A flat ribbon cable connecting a camera directly to a board. Lower latency than USB.
DOFDegrees of Freedom. How many independent axes a robot joint system can move.
PLA / PETGCommon 3D printing plastics. PLA cheaper and easier; PETG tougher and more heat-resistant.
Edge AIRunning AI models on local hardware instead of cloud servers. Low latency, private, no API costs.
FSRForce Sensitive Resistor. A thin pad that changes electrical resistance when pressed. Used in gripper fingertips.
ESCElectronic Speed Controller. Sits between a compute board and a brushless motor.

2. The four columns

Every robot body is made of four types of components: sensors (eyes/ears/skin), compute (brain), actuators (muscles), structure (skeleton). The relationship is directional: sensors perceive, compute decides, actuators act, structure holds it all.

The four hardware columns of a Physical AI robot Four hardware components compose every robot: sensors collect information, compute processes it, actuators act on the world, structure holds everything together. The four columns of robot hardware Sensors Eyes, ears, skin RGB, depth, LIDAR Audio, IMU, force perceives Compute Brain hardware Jetson family CUDA + TOPS decides Actuators Muscles Servos, DC, steppers Grippers, rotors acts Structure Skeleton PLA, PETG, resin Aluminum, carbon holds it all Sensors drive compute requirements; actuators barely need compute. A servo command is microseconds; processing a camera image is millions of operations.
Every robot is made of the same four columns. The system type (arm, AMR, humanoid, drone) is a different combination at different scales.

3. Sensors

How a robot perceives the world. Different sensors capture different information. The more and better sensors you add, the more compute power you need to process the data in real time.

Vision: RGB cameras

Two connection types. USB (UVC standard): plug into any USB port on any computer; works immediately with no drivers; good for learning. CSI cameras: flat ribbon cable directly into a Jetson or Raspberry Pi; lower latency than USB; used when speed matters.

ProductTypeResolutionPrice
Logitech C920 / C922USB1080p~$50-70
innomaker UVC CameraUSB1080p, 130° wide~$25-35
Raspberry Pi Camera Module 3CSI12MP~$25
Arducam (various models)CSI / USBVarious$15-80

Vision: depth cameras

Two technologies: stereo vision (two cameras side by side, like human eyes; works outdoors; RealSense D435, ZED) and time-of-flight (emits infrared light and measures bounce; more accurate at close range; struggles in bright sun).

ProductRangeBest forPrice
Intel RealSense D4350.2-10mGeneral robotics, wide FOV$314
Intel RealSense D435i0.2-10mD435 + built-in IMU for motion tracking$334
Intel RealSense D4050.07-0.5mClose-range, gripper-mounted$272
Intel RealSense D4550.4-6mLonger range, outdoor capable$419
Stereolabs ZED 2i0.2-20mWide angle, IP66, AI-powered~$450+
Stereolabs ZED X Nano0.03m+Wrist-mount, sub-mm accuracyPre-order

Vision: LIDAR

Shoots thousands of laser beams per second in a spinning pattern, builds a "point cloud" 3D map. 2D LIDAR scans a single horizontal plane (good for indoor navigation, $99-600). 3D LIDAR scans multiple planes ($500-75K+, used in self-driving cars).

ProductTypeRangePrice
Slamtec RPLIDAR A12D, 360°12m$99
Slamtec RPLIDAR A22D, 360°18m$199-229
Slamtec RPLIDAR A32D, 360°25m$599
Unitree 4D LIDAR L13DLong rangeBundled with Go2

Audio: microphones and mic arrays

A single USB microphone captures voice commands. A microphone array detects direction and cancels background noise. ReSpeaker Mic Array v2.0 (~$60-80) is the standard for far-field voice on Jetson. Running Whisper (speech-to-text AI) locally on Jetson is medium-compute.

Motion: IMU and encoders

IMU (Inertial Measurement Unit): a small chip with accelerometers and gyroscopes. Tells the robot which way is up, if it is tilting, how fast it is moving. Essential for balancing humanoids and stabilizing drones. Often built into compute boards or depth cameras (the RealSense D435i includes one).

Encoders: built into servo motors. Report the exact angle of each joint back to the compute board. The SO-ARM101's STS3215 servos have 12-bit magnetic encoders (4,096 positions per rotation).

Touch: force sensors

FSR pads (Force Sensitive Resistors): thin flexible pads that change electrical resistance when pressed. Attached to gripper fingertips, they let the robot feel how hard it is squeezing. Without them, a robot cannot tell if it is about to crush a grape or drop a heavy mug. $10-50 per pad. Tactile sensor arrays are the more advanced version with a grid of sensing points; still mostly research.

Not in this list (specialized sensors)

Thermal / infrared cameras (heat vision), radar (used in cars alongside cameras and LIDAR), ultrasonic sensors (cheap parking-style distance), gas / chemical sensors, GPS modules, barometric pressure. Real but not needed for learning and first builds.

4. Compute

Compute is where the AI brain runs. It takes sensor data in, processes it through AI models, and sends commands to actuators. The compute board determines how many sensors run simultaneously and how complex the AI models can be.

The NVIDIA Jetson family (the dominant platform)

All Jetson boards run the same software (JetPack SDK). Code written on one works on any other. JetPack 6.2 (current production release) includes CUDA 12.6, TensorRT 10.3, cuDNN 9.3, Jetson Linux 36.4.3 (Linux Kernel 5.15, Ubuntu 22.04).

BoardAI powerMemoryHandlesPrice
Jetson Orin Nano Super (recommended)67 TOPS8GB2-4 cameras, local LLMs (8B at ~15 tok/s), real-time object detection$249
Jetson Orin NX100 TOPS8-16GBMulti-model inference, complex robotics$400-600
Jetson AGX Orin275 TOPS32-64GBMultiple large models simultaneously, commercial robots$1-2K
Jetson Thor2,070 FP4 TFLOPS128GBFoundation models (GR00T) on-device, humanoid robots$1,999+
Jetson T40001,200 FP4 TFLOPS64GBBlackwell in smaller module. Announced CES 2026.$1,999 (volume)

What TOPS means in practice: the Orin Nano's 67 TOPS can process 2-4 camera streams at 30 fps while simultaneously running an object detection model (YOLO) and a local LLM (Llama 8B). Enough for a robotic arm, a security camera system, or a voice assistant. Step up the ladder when you need more.

Buy: Jetson Store | Orin Nano Super Dev Kit

Alternatives

BoardAI powerBest forPrice
MacBook Pro (Apple Silicon)~38 TOPS Neural EngineDevelopment, testing, Ollama. Not a permanent robot brain.Already owned
Raspberry Pi 5 + Hailo AI HAT+13 TOPSSingle camera projects, entry-level edge AI~$120
Google Coral USB Accelerator4 TOPSAdding AI to any USB computer (TF Lite only)~$60
IGX Platform (NVIDIA)Same as Jetson ThorIndustrial / safety-certified (ISO 13849). Factories, hospitals.Enterprise
Why NVIDIA dominates
CUDA compatibility. Nearly every AI framework (PyTorch, TensorFlow, YOLO, Whisper) runs on NVIDIA GPUs. A Jetson board runs the same CUDA code as a $30,000 data center GPU. Code and skills transfer up and down the ladder without rewriting. No other edge AI platform has this advantage.

5. Actuators

How a robot acts on the world. Every physical movement (rotating a joint, spinning a wheel, opening a gripper, generating lift) is performed by an actuator. Actuators are surprisingly light on compute; the AI decision-making is the heavy part. Sending "rotate to 45 degrees" to a servo takes microseconds.

Servo motors (precise angle rotation)

Receives a command like "rotate to 135 degrees" and moves to that exact position. Holds against external force. Built-in encoder reports the current angle back to compute, creating a feedback loop. Servos can be daisy-chained: each motor has a unique ID, all share the same data cable. The SO-ARM101 daisy-chains 6 servos per arm.

ProductTorqueVoltagePrice eachUsed in
Feetech STS321519.5 kg/cm (7.4V) or 30 kg/cm (12V)7.4V or 12V$15-29SO-ARM100 / 101
ROBOTIS Dynamixel AX-12A15 kg/cm12V~$50Entry-level humanoid kits
ROBOTIS Dynamixel XL43021 kg/cm12V~$50Replacement for AX-12A
ROBOTIS Dynamixel XM43043 kg/cm12V~$115Research arms, professional kits

Feetech vs Dynamixel: Feetech STS3215 costs $15-29 with open-source Python control. Dynamixel XM430 costs $115 with more torque and a professional SDK. The SO-ARM101 uses Feetech because the price / performance ratio is excellent for learning. Dynamixel is the industry standard for professional and research robots.

Feetech | ROBOTIS Dynamixel | WowRobo (STS3215)

Other actuator types

TypeWhat it doesUsed inPrice range
DC motorsSpin continuously at variable speed. Speed controlled by voltage.Wheels (JetBot, JetRacer), conveyor belts, fans$5-30 each
Stepper motorsRotate in exact discrete steps (e.g., 1.8° per step). Precise without feedback sensors.3D printers, CNC machines, camera gimbals$10-50 each
Linear actuatorsPush / pull in a straight line. Like a motorized piston.Industrial arms, adjustable platforms, lifts$20-200
GrippersEnd effectors that open and close. Parallel jaw (two flat plates) or adaptive (conforms to shape).End of any robotic arm. SO-ARM101 includes one.$20-500
Rotors / propellersSpin at high RPM to generate lift. Quadcopters use 4 rotors.Drones, quadcopters, fixed-wing UAVs$10-50 each + ESC
SpeakersSound is an actuator too. A robot that can talk back, beep, or play alerts.Voice assistants, companion robots, Reachy Mini$10-50

Not covered: pneumatic actuators (compressed-air, factory robots), hydraulic (fluid-powered, heavy construction), shape-memory alloys (experimental soft robotics), piezoelectric (tiny precise movements in surgical tools).

6. Structure

The physical frame that holds sensors, compute, and actuators together and gives the robot its form factor. Material depends on scale, environment, and budget.

MaterialWhat it isUsed inCostNotes
3D printed (PLA)Polylactic Acid. Cheapest, most common 3D printing plastic. Biodegradable.SO-ARM101 frames, custom mounts, prototypes$5-50Easy to print. Brittle under stress. Not heat-resistant.
3D printed (PETG)Polyethylene Terephthalate Glycol. Tougher than PLA.Parts needing more durability than PLA$10-60More flexible, heat-resistant. Harder to print.
3D printed (resin)Photosensitive resin cured by UV light. Smoother surfaces, finer detail.Waveshare SO-ARM kits use resin parts$30-80More brittle than PLA but better tolerances.
Aluminum extrusionT-slot aluminum profiles that bolt together like erector sets.JetBot / JetRacer frames, workbenches, custom rigs$20-100Strong, reusable, easy to modify. Standard sizes: 2020, 2040.
Metal chassisSteel, aluminum, or titanium sheet / machined parts.Vehicles (Tesla), industrial arms (FANUC, ABB)$100-10K+Production-grade. Too expensive for learning.
Carbon fiberExtremely light and rigid composite.Drones (weight is critical for flight time), racing$50-500Every gram matters when something flies.
Injection molded plasticPlastic parts made in mass-production molds.Consumer products (Reachy Mini body, phone cases)N/A for hobbyExpensive molds, cheap per unit. Not for prototyping.

7. How sensors connect to compute

This is the key hardware relationship. Actuators barely need compute power. Sensors drive the compute requirements because processing camera images and LIDAR point clouds takes millions of operations per frame.

Sensor loadExample setupMinimum compute
Light1 USB camera + motor encodersAny (Mac, Raspberry Pi, or Jetson)
Medium2 cameras + 1 depth sensor (RealSense)Jetson Orin Nano (67 TOPS)
Medium4 cameras at 1080p 30fps (security system)Jetson Orin Nano to NX
HeavyLIDAR + cameras + IMU (autonomous navigation)Jetson Orin NX or AGX Orin
HeavyFull humanoid sensor suite (all types)Jetson Thor
For a first arm build (SO-ARM101 + Jetson Orin Nano)
1-2 cameras (light to medium compute), 12 servo motor encoders (light), optionally a RealSense depth camera (medium). The Orin Nano handles all of this with room to spare.

8. Manufacturer directory

Where to actually buy. Categorized by what they sell.

Compute boards

Sensors

Actuators (motors and servos)

Kits and arms