Hardware — the parts you can actually buy
Every algorithm on this site eventually runs on a real device with a real price, mass, power draw and error budget. This section is a working survey of the sensors and compute platforms available on the market: what each technology can physically do, how the grades differ, what the datasheet numbers mean once you put them in a filter, and what it costs you in size, weight, power and dollars.
Symbols used on this page are defined in the Nomenclature → inertial navigation.
Specifications are collected from public datasheets and are indicative rather than contractual — always confirm against the current datasheet and your own bench tests. Bench results of our own will be added here as they are run.
Inertial & attitude sensing
The sensors that tell you how you moved and where you are pointed — the parts whose error terms dominate every navigation budget on this site.
IMUs
Consumer MEMS to strategic grade: bias stability, ARW, VRW, scale factor and what each grade buys you in free-inertial drift. Real part numbers with real numbers.
HG9900 tutorial: increments → EKF
A build-it walkthrough for a navigation-grade RLG unit: why it publishes integrals instead of rates, why the angle arrives in two halves, and the complete error-state filter math — with runnable code.
Gyroscopes
How the sensing physics differ — vibratory MEMS, fibre-optic, ring-laser, hemispherical resonator — and why that sets the noise floor, the g-sensitivity and the price.
Star trackers
Cross-boresight vs roll accuracy, update rate, exclusion angles, lost-in-space capability, and the CubeSat-to-flagship range of options.
Ranging & positioning
The sensors that measure the world outside the vehicle, and the receiver that ties you to an absolute frame.
LiDAR
Spinning vs solid-state vs FMCW, 905 nm vs 1550 nm, range at reflectivity, points per second, and what the choice does to a perception stack downstream.
GNSS receivers
Single vs multi-frequency, RTK and PPP, anti-jam and anti-spoof features, timing holdover, and the interfaces that matter for tight coupling.
Compute
Where the perception models actually have to run.
Edge compute (Jetson / Orin)
TOPS vs achievable FPS, memory bandwidth as the real limit, power modes, and what actually fits a BEVFusion-class model.
Measured: model size, FPS & GPUs →
The companion benchmark page from the perception study — the same boards under a real model, with latency percentiles instead of headline TOPS.
How to read a sensor datasheet
Datasheets are written to be flattering. A few habits make them comparable:
- Find the integration time. Noise specs are meaningless without one. An Allan-variance curve tells you which error dominates at which averaging time — white noise falls as $1/\sqrt\tau$, bias instability flattens out, and random walk grows. Match the curve's minimum to your update interval.
- Separate repeatability from stability. Turn-on repeatability is how much the bias changes between power cycles (you must estimate it every time); in-run stability is how much it wanders once running (that sets your FOGM time constant).
- Check the temperature qualifier. "0.01°/hr" at constant 25 °C and "0.01°/hr over −40…+85 °C" are different products at different prices.
- Look for the axis definition and the g-sensitivity. Vibration-rectification and g-dependent bias are what actually ruin MEMS performance on a vehicle.
- Interface and latency are error terms too. An excellent sensor with an unknown, jittery time-tag becomes a mediocre one — see the time-tagging discussion.
The trade space at a glance
| Class | Gyro bias stability | ARW | Free-inertial drift | Typical use |
|---|---|---|---|---|
| Consumer MEMS | 100–1000 °/hr | 1–3 °/√hr | km in seconds–minutes | phones, drones with GNSS |
| Industrial MEMS | 3–30 °/hr | 0.2–1 °/√hr | ~km in minutes | AGVs, tightly-coupled GNSS/INS |
| Tactical | 0.5–3 °/hr | 0.05–0.2 °/√hr | ~1 nmi in ~10 min | guided munitions, UAVs |
| Navigation | 0.005–0.05 °/hr | 0.001–0.01 °/√hr | ~1 nmi/hr | aircraft, launch vehicles |
| Strategic | < 0.001 °/hr | < 0.0002 °/√hr | < 0.1 nmi/hr | submarines, ICBMs |
Detailed per-part tables live on the IMU and gyroscope pages; the theory behind these numbers is on the IMU & inertial navigation page.
Planned bench tests
Specs are a starting point; the interesting part is measuring them. The intent for this section is to add reproducible bench comparisons:
- Allan variance from multi-hour static logs, to extract ARW / bias instability / rate ramp per unit and check them against the datasheet.
- Turn-on repeatability across many power cycles and temperatures.
- Static drift races — free-inertial position error vs time for each IMU grade, side by side.
- LiDAR range/reflectivity curves and point-density-vs-range measurements on a fixed target board.
- Inference benchmarks — the same detector exported to TensorRT across compute boards, reporting achieved FPS, latency percentiles and power, not just TOPS.
Related: IMU theory · star-tracker theory · GNSS theory · the perception stack these sensors feed.