robotics//sensor

A sensor is a device that converts a physical quantity (temperature, acceleration, light, pressure) into one a circuit can read, almost always a voltage, and it is the only way anything the world does enters a control loop, an estimator or a model. Each **transduction principle** brings its own virtues and vices. The MEMS accelerometer of a drone is a microscopic mass on silicon springs whose displacement changes a capacitance; a strain gauge changes resistance when stretched; a thermocouple produces microvolts from a temperature difference. Small resistance changes are read in a Wheatstone bridge, four resistors whose output is zero when balanced and a few millivolts per volt of supply when one of them moves, which rejects the large common part and keeps only the change.

A sensor is a device that converts a physical quantity (temperature, acceleration, light, pressure) into one a circuit can read, almost always a voltage, and it is the only way anything the world does enters a control loop, an estimator or a model. Each transduction principle brings its own virtues and vices. The MEMS accelerometer of a drone is a microscopic mass on silicon springs whose displacement changes a capacitance; a strain gauge changes resistance when stretched; a thermocouple produces microvolts from a temperature difference. Small resistance changes are read in a Wheatstone bridge, four resistors whose output is zero when balanced and a few millivolts per volt of supply when one of them moves, which rejects the large common part and keeps only the change.

Every real sensor lies a little: with noise, with a bias, with a delay, with the wrong time on its reading. The sensor error model writes all of that in one line, and most of the work with sensors is knowing those terms for the part in hand.

Navigation sensors measure the vehicle itself. The IMU gives acceleration and rotation rate at kilohertz and drifts; the GNSS receiver gives position slowly but without drift; the barometer gives height and wanders with the weather; the magnetometer gives heading and is disturbed by every motor and steel beam; the rotary encoder gives a shaft angle per edge. None measures exactly what the vehicle needs, which is why they are fused.

Perception sensors measure the surroundings: the camera, the lidar and the radar, each failing in different weather and light (perception). Process and force sensors measure a plant: the temperature sensor, the strain gauge in a load cell, the transmitters on a 4-20 mA loop.

Between the physics and the code sits a measurement chain of conditioning, conversion, bus and software, each link with its errors; a smart sensor carries half that chain inside the chip. What a sensor cannot measure can sometimes be reconstructed from others and a model, a virtual sensor. Before any of them is trusted it is calibrated (sensor calibration) and its valid conditions are known (operating envelope).

Mounting matters as much as calibration. A vibration accelerometer is bolted on with a stud to keep its high frequencies, the IMU of a flight controller is isolated from them on dampers, and the magnetometer goes far from power cables, often on the GPS mast.

Packaged sensors can resemble chips, but their package often exposes a physical interface (a port, membrane, optical window, proof mass or electrode), and their performance depends more on materials, packaging, calibration and analog design than on the newest semiconductor node. The output reaches an MCU or SoC through an ADC, MMIO or a serial bus; in a smartphone the sensors report to an MCU, which decides what to pass on to the MPU. An ASIC on the sensor can filter or threshold at minimal power, while an MCU running an FFT or TinyML can be updated.