Sensors, computers, actuators, and power systems explained through physical principles, specifications, real products, selection tradeoffs, and limitations.
This is an editorially selected reading order. Begin at step 1 or go directly to the topic you need.
We compare the embedded computers that serve as a robot's "brain" across three design philosophies — the general-purpose SoC of the Raspberry Pi 5, the GPU-integrated AI computer of NVIDIA Jetson Orin, and the dedicated ASIC of Google Coral. From what the TOPS/W efficiency metric actually means, to why the newbot project chose a general-purpose SoC, to the real 2026 story of an AI-datacenter DRAM crunch hitting the hobbyist SBC market.
LiDAR measures distance from the round-trip time of light, from the rotating Velodyne Puck to the non-rotating Livox Mid-360S that drove costs down, to the Chinese players Hesai and RoboSense. We compare real specs from actual products in production — and trace the industry's real backstory, from the DARPA Grand Challenge to the DJI spinoff to the collapse of Volvo's deal with Luminar.
A monocular camera reads lane lines, signs, and obstacles through a single lens, gaining low cost and a small footprint at the price of a fundamental limitation — a single image alone cannot fix absolute distance. We look at real examples from Mobileye EyeQ6L, Tesla HW4, Continental MFC500, and comma.ai to see how each works around that constraint, and trace how a Hebrew University lab became Mobileye and was acquired by Intel for over $15 billion, plus Tesla's 2021 decision to "delete" radar from Tesla Vision.
A stereo camera triangulates distance from the disparity between two lenses, solving the absolute-scale problem that fundamentally limits monocular cameras. We look at real specs from the Stereolabs ZED 2i, Luxonis OAK-D, and e-con Systems TARA, and trace the baseline-length tradeoff through the evolution of Subaru's EyeSight, which has shipped on over 5 million vehicles since 2008 — plus the story of how EyeSight's creator left Subaru to found his own stereo-vision startup.
Active stereo projects a pattern with an infrared emitter and does stereo matching; structured light reads depth from how a known pattern distorts; time-of-flight reads distance directly from a phase shift — three coexisting approaches to depth cameras. We look at the principles behind each and real examples from the Intel RealSense D435i, Microsoft Kinect, Orbbec Femto Bolt, and iPhone TrueDepth, and trace how Israeli startup PrimeSense's technology gave birth to Kinect, got acquired by Apple, and was reborn as Face ID.
An IMU pairs an accelerometer with a gyroscope to directly sense a vehicle's own translational acceleration and rotational rate. Unlike cameras or LiDAR, it depends on nothing external — but that self-containment comes with an inescapable fate: error accumulates over time (drift). We work through the MEMS gyroscope principle, Allan-variance noise analysis, the grading scale from consumer chips to navigation-grade modules, and Xsens's MTi series product strategy.
We compare the two operating principles behind the rotary encoder — optical, which reads position by interrupting light, and magnetic, which reads the direction of a magnetic field — from a motor's shaft angle and speed sensor. From the encoder industry's surprising origin at an organ maker, to the newbot project's real measurement of a 1.2m straight-line run being misread as an 80° turn when the actual rotation was 6.5°, we trace the real story of the encoder that underpins — and sometimes betrays — wheel odometry accuracy.
RTK combines satellite positioning with correction data from a base station, shrinking the several-meter error of standalone GNSS positioning down to centimeter level. We work through how dual-frequency positioning cancels ionospheric delay, how u-blox's ZED-F9P drove low-cost RTK into the mainstream, Trimble RTX and Japan's own Michibiki (QZSS), and the real-world example of Kubota's nationwide network of 342 RTK base stations for farm machinery — and how each of these changes what positioning accuracy you actually get.
The magnetometer, which senses the direction of Earth's magnetic field to derive heading, is the one sensor that keeps returning an absolute-direction external reference as long as it has power — from a smartphone's electronic compass, to yaw-angle correction on a drone, to attitude control on a CubeSat. This article lays out the four underlying physical principles — the Hall effect, AMR, GMR, and fluxgate — compares real chip specs across the Bosch BMM150, STMicroelectronics LIS3MDL, and Honeywell HMC5883L, covers hard-iron/soft-iron error and its calibration, AHRS sensor fusion with an IMU, and the technology's history from ore prospecting through WWII submarine hunting and lunar exploration to today's CubeSat attitude control.
Millimeter-wave radar measures distance and relative velocity from the reflection of radio waves, making it the sensor most resistant to rain, fog, glare, and other adverse conditions that give optical sensors trouble. We work through how FMCW's beat frequency yields distance and velocity simultaneously, radar's technology lineage running back to Chain Home in the Second World War, Tesla's abandonment and later reinstatement of radar, the 4D imaging specs of the Continental ARS540, and the real-world example of Honda SENSING 360's five-radar 360-degree coverage — and how each of these changes the range/field-of-view/resolution trade-off.
An ultrasonic sensor measures a nearby obstacle's distance from the round-trip time of an ultrasonic pulse — the acoustic counterpart to LiDAR. It can see straight through what trips up optical sensors, transparent glass and black objects included, but its effective range tops out at a few meters. We work through bat-echolocation research, sonar's development in the First World War, the Bosch USS Gen 6's measured range, the Murata MA40S4S and MaxBotix modules that are staples of robotics, and the real-world example of Toyota's eight-sensor Intelligent Clearance Sonar — and how each of these changes what performance you actually get.
Unlike conventional cameras that keep capturing frames (still images) at fixed intervals, event cameras (Dynamic Vision Sensor, DVS) have each pixel independently report only the instant its brightness changes, at microsecond resolution. We trace the real chips — Prophesee Metavision, iniVation DVXplorer/DAVIS, Samsung DVS-Gen4 — the neuromorphic-engineering lineage from 1988's silicon retina research to iniVation's 2024 acquisition, the stereo-event-camera autonomous-driving dataset DSEC, and SLAM research running onboard UAVs.
Every temperature sensor does the same one thing — turns heat into an electrical signal — but the physics underneath splits into four separate families: the Seebeck effect in thermocouples, the resistance-temperature relationship in RTDs and thermistors, the bandgap behavior of semiconductor ICs, and the Stefan-Boltzmann law behind infrared radiation thermometers. This piece traces the history from Seebeck's 1821 misunderstanding, through the Callendar-Van Dusen equation that calibrates PT100 sensors, to a spec comparison of real ICs like the DS18B20, TMP117, and MLX90614, and a real-world case of an F1 car reading tire temperature with 12 non-contact IR sensors.
Force sensors give a robot arm its 'sense of the arm,' and tactile sensors give a robot hand its 'sense of the fingertips.' We trace the physics from strain gauges through three tactile-sensing approaches — piezoresistive, capacitive, and optical — and survey real products from ATI, Bota Systems, and XELA Robotics to Shadow Robot's prosthetic-style hand carrying 129 sensors, alongside the true story of the strain gauge's simultaneous invention in 1938.
We compare the actuators that serve as a robot's "muscles" across three angles — the BLDC motor, the stepper motor, and the servo motor. These three aren't actually a parallel classification; motor construction (BLDC/stepper) and control architecture (servo) are two different axes that get conflated. Untangling that confusion, we trace the real torque/response/control tradeoffs from NASA's brush-wear-in-vacuum problem to the birth of Wiener's cybernetics.
What determines a robot's or drone's range and safety isn't the motor or the sensor suite — it's the battery and its BMS (Battery Management System). From the electrochemistry of lithium-ion cells to cell balancing and protection circuits, we compare real specs from Molicel's high-discharge cells, DJI's intelligent flight batteries, and Samsung SDI's push into next-generation solid-state batteries — and look at what the Boeing 787 fire taught the industry about the limits of a BMS.
Relate motor load changes to voltage drops and locate faults across the battery, wiring, converter and computer.
What carries a sensor's measured value outside its enclosure is, almost always, a radio wave. We trace how Wi-Fi's OFDM uses orthogonal subcarriers to overcome multipath, how BLE's GFSK modulation and star topology enable duty-cycled operation, and how LoRa's Chirp Spread Spectrum (CSS) trades range for data rate across the SF7-12 spreading-factor range — all through the real numbers of chips like the ESP32-C6, Nordic nRF52840, Semtech SX1262, TI CC2652R7, and MultiTech mDot. From the 1999 naming drama behind Wi-Fi, to the name Bluetooth born from a 10th-century Danish king's epithet in 1996, to the story of how LoRa, built by a small French startup, was acquired by Semtech, to the real-world example of ZENNER's LoRaWAN smart-meter network scaling past 10 million devices, we look at how to choose frequency band, transmit power, sensitivity, and topology.
Rather than comparing whole SBCs like the Raspberry Pi or Jetson, this piece narrows in on the "AI inference accelerator" chips that sit on top of them, comparing three design philosophies — Google Coral (Edge TPU), Intel Movidius (Myriad X), and NVIDIA Jetson Orin NX. It traces the principle behind why INT8 quantization trades accuracy for speed, through a real example of a Skydio drone running nine neural networks concurrently on a single Jetson TX2, to the state of Coral, effectively abandoned by Google but kept alive by its community.