Sensor stack reference

Sensors: Cameras, LiDAR, Radar, IMU, GPS, and Sensor Fusion

The practical question is not “which sensor is best?” It is “which combination survives the environment, keeps timing honest, and gives the estimator enough truth to correct the rest.”

Built for robotics, autonomy, drones, phones, AR, and field systems where you need geometry, motion, time, and position to agree.

Sensor fusion cockpit diagram Camera, LiDAR, radar, IMU, GNSS, and thermal layers around a robot on a calibrated grid. camera LiDAR radar GNSS IMU + time

Quick Reference: what each sensor really buys you

Sensor Measures Best at Weak at Typical failure mode Use it when
Camera Radiance projected into pixels; can infer color, texture, labels, and lanes. Semantics, classification, OCR, cheap dense detail. Absolute depth, darkness, glare, weather, motion blur. Rolling-shutter wobble, exposure washout, occlusion. You need “what is it?” more than “exactly where is it?”
LiDAR Time of flight to build a point cloud. Geometry, shape, obstacle outlines, 3D mapping. Color, texture, fog/rain/dust scattering, black absorbers. Ghost points, dropouts, multipath off shiny surfaces. You need precise 3D structure in controlled or semi-controlled air.
Radar Range, radial velocity, and sometimes angle from RF reflections. Motion, weather tolerance, long-range detection. Fine shape, semantic detail, clean separation in dense clutter. Multipath, clutter, poor angular resolution versus vision. You care about moving objects and the weather is bad.
IMU Acceleration and angular rate. Fast local motion, stabilization, prediction between slower sensors. Absolute position and long-term heading without correction. Bias drift, integration drift, vibration sensitivity. You need continuity at high rate, then correction from other sensors.
GNSS / GPS Global position, velocity, and time from satellites. Outdoor absolute position, map anchoring, time sync. Indoors, urban canyons, foliage, tunnels, intentional interference. Multipath, jamming, spoofing, satellite masking. You need global context and can see the sky.
Ultrasonic Near-field distance from sound reflection. Parking, short-range proximity, tank or bin level. Soft surfaces, long range, precise lateral detail. Angle sensitivity, wind, temperature variation, dead zones. Simple close-range detection is enough.
Thermal Infrared emission mapped to apparent temperature. Night work, heat leaks, people or equipment contrasts. Fine texture, exact material identity, internal temperature. Emissivity error, reflective surfaces, ambient wash. Visible light is poor but heat contrast is useful.
Wheel odometry Relative motion from encoder counts and kinematics. Short-horizon motion on flat or structured ground. Slip, bumps, stairs, loose terrain. Dead-reckoning drift that accumulates with distance. You need local pose prediction and can tolerate drift correction.

Sensor mental model

Measurement vs inference

A sensor reports a measurable physical effect; the system infers state from it. A camera measures light, then software infers a lane, a face, or a barcode. A GNSS receiver measures satellite timing and then infers position and time.

Example: a 12.5 MP Basler camera can see a pallet label; it does not directly tell you the pallet is 2.6 m away unless you add geometry or another depth source.

Bias, drift, noise

Noise is the short-term wobble; bias is the persistent offset; drift is bias integrated over time. IMUs live or die on this distinction because a tiny gyro bias becomes a large angle error after minutes.

Example: Analog Devices notes gyro drift comes from bias instability and angular random walk; that is why you never leave an IMU uncorrected for long runs.

Calibration and time

Extrinsics tell you where sensors sit relative to one another; timing tells you whether they describe the same instant. A perfect camera frame and a perfect LiDAR scan are still wrong if they are 120 ms apart.

Example: Ouster lists time synchronization via gPTP and PTP, because multi-sensor systems need aligned timestamps more than they need raw throughput.

Representative hardware snapshot

Example Snapshot What it teaches
Basler ace 2 a2A3536-9gcPRO 12.5 MP, 9 fps, Sony IMX676 CMOS sensor, net list price €419.00 on the Basler shop as of July 2026. Camera choice is about pixel size, shutter behavior, interface, and exposure strategy, not just megapixels.
Ouster OS0 Ultra-wide 90° FOV, 35 m range at 10% target, 0.25 cm precision up to, 5–40 Hz, IP68/IP69K. Short-range LiDAR wins when you need dense near-field geometry around a robot or warehouse vehicle.
Ouster OS1 Max 45° FOV, 200 m range at 10% target, 0.25 cm precision up to, 15–25 W, 128 or 256 channel options. Longer range usually costs power and packaging complexity, but it buys more stand-off for faster vehicles.
TI AWR2944EVM 76–81 GHz mmWave radar evaluation module with 4TX/4RX antenna; TI lists it at $628.95. Automotive radar is a range-and-velocity sensor first; the point cloud is a byproduct of signal processing.
Bosch BMI270 6-axis IMU, ±2g to ±16g accel, ±125 dps to ±2000 dps gyro, up to 6.4 kHz gyro ODR. IMUs are fast enough to run prediction loops, but they need correction from something that does not drift the same way.
u-blox ZED-F9P Multi-band GNSS with centimeter-level accuracy in seconds with RTK, fast convergence, and multi-constellation reception. GNSS goes from “good enough for navigation” to “survey-grade” only when corrections and sky view are good.
MaxBotix MB1040 6 to 254 inches, 20 Hz read rate, narrowest detection field, $23.65. Cheap ultrasonic sensors are excellent for short-range proximity, but they are not a general-purpose depth camera.
FLIR C8 320 × 240 thermal resolution, compact thermal camera, $899 on the official FLIR store. Thermal imaging is about heat contrast and inspection utility, not seeing through walls or measuring internal temperature directly.

Sensor deep dives

Camerassemantics first

Definition: a camera samples reflected light and turns it into pixels that software can classify, measure, or segment. Use it when “what is it?” matters more than direct depth.

Example: Basler’s ace 2 a2A3536-9gcPRO exposes a Sony IMX676 CMOS sensor, 12.5 MP resolution, and 9 fps. Sony’s industrial pages split camera behavior into rolling-shutter and global-shutter modes.

Gotcha: rolling shutter distorts fast motion and cheap exposure control can wash out glare or shadow. Do not treat camera AI output as a measurement of distance unless you have stereo, structure-from-motion, or another depth source.

LiDARgeometry first

Definition: LiDAR emits laser pulses and measures return time to produce a 3D point cloud. It is the cleanest path to geometry when you need actual shape, not just texture.

Example: Ouster’s OS0 is a 90° ultra-wide sensor for short-range robotics, while OS1 Max reaches 200 m at 10% target with a 45° field of view. That is the trade: near-field breadth versus long-range reach.

Gotcha: fog, rain, dust, black absorptive surfaces, and reflective clutter can all degrade returns. If your environment is dirty or highly reflective, LiDAR is usually a piece of the stack, not the whole stack.

Radarmotion in weather

Definition: radar sends RF energy and reads the reflection to estimate range and radial velocity; modern mmWave systems also infer angle and object clusters.

Example: TI’s AWR2944EVM operates in the 76 to 81 GHz band and uses a 4TX/4RX antenna. TI lists the board at $628.95, which is a useful reminder that useful radar is still not toy-priced.

Gotcha: radar sees motion and multipath better than shape. Dense urban clutter, metal reflectors, and near-field interference can create ghosts or ambiguous returns.

IMUprediction engine

Definition: an IMU combines accelerometers and gyroscopes to measure linear acceleration and angular rate. It gives the estimator high-rate continuity when the slower sensors are between samples.

Example: Bosch’s BMI270 is a 6-axis IMU with ±2g/±4g/±8g/±16g accel ranges and ±125 dps through ±2000 dps gyro ranges. That is plenty for stabilization, but it still drifts.

Gotcha: gyro drift grows because of bias instability and angular random walk. Without correction from vision, GNSS, magnetometer, or wheel/scene constraints, dead reckoning will walk away from reality.

GNSS / GPSglobal context

Definition: GNSS receivers use satellite timing to estimate position, velocity, and time. GPS.gov describes GPS as a U.S.-owned PNT utility, and the system becomes most useful when you can also trust the sky view and spectrum.

Example: u-blox’s ZED-F9P module advertises centimeter-level accuracy in seconds with multi-band RTK. In practice that means a rover or vehicle can anchor itself globally instead of merely locally.

Gotcha: indoor use is usually a non-starter, and even outdoors you must worry about multipath, masking, jamming, and spoofing. GPS.gov and FAA guidance explicitly call out interference, because it is not an edge case anymore.

Ultrasoniccheap near-field

Definition: ultrasonic sensors measure distance by timing sound reflections. They are a low-cost solution for close-range obstacle detection and level sensing.

Example: MaxBotix’s MB1040 LV-MaxSonar-EZ4 ranges from 6 inches to 254 inches, reads at 20 Hz, and lists at $23.65. That is why it appears in parking, bin, and small-robot designs.

Gotcha: the beam is wide, soft targets absorb energy, and the environment changes the reading. Use it for short-range proximity, not for precise mapping.

Thermalheat contrast

Definition: thermal cameras detect infrared energy and turn it into an image, often calibrated for temperature inspection. They help when visible light is weak or misleading.

Example: FLIR’s official store lists a 320 × 240 compact thermal camera for $899, and the E54 reports temperatures up to 650°C. That is enough for inspections, not for magical X-ray vision.

Gotcha: emissivity, reflections, and surface coatings can mislead the reading. Thermal sees surface temperature, not hidden internal structure.

Wheel odometrylocal motion only

Definition: wheel odometry estimates motion from encoder counts and robot kinematics. It is the simplest relative-motion signal in ground robots.

Example: in ROS state estimation, the projected pose is corrected by perceived sensor data, which is exactly why odometry is useful as a prediction input and dangerous as a final answer.

Gotcha: wheel slip, bumps, stairs, and loose terrain accumulate error fast. Treat odometry as a short-horizon bridge, not a map of the world.

Stack decisions: what to use when

Use case Recommended stack Why this stack works What to add next
Phone AR / consumer vision Camera + IMU + optional depth or GNSS Camera handles semantics, IMU gives motion continuity, depth or GNSS adds global scale when available. Add LiDAR or time-of-flight if you need better occlusion handling or room-scale geometry.
Warehouse AMR Camera + LiDAR + IMU + wheel odometry LiDAR gives floor and obstacle geometry, camera reads labels and signs, IMU/odometry stabilize motion between scans. Add GNSS only if the building or yard has usable sky view; otherwise build off map-based localization.
Outdoor drone IMU + GNSS + camera, optionally barometer and LiDAR IMU handles fast motion, GNSS anchors global position, camera supports landing and obstacle perception. Add visual odometry or LiDAR for hover and GNSS-denied resilience.
Highway vehicle Camera + radar + LiDAR + IMU + GNSS Camera classifies lanes and signs, radar tracks motion in bad weather, LiDAR improves geometry, IMU smooths everything. Use redundancy for safety-critical features; do not lean on one modality for every condition.
Precision agriculture GNSS RTK + IMU + camera + optional LiDAR/radar RTK gives centimeter-level positioning, IMU stabilizes turns, camera handles row detection and crop state. Add radar if dust, spray, or foliage make vision unstable.
Night inspection / maintenance Thermal + camera + IMU + GNSS Thermal exposes heat anomalies, camera preserves context, IMU/GNSS preserve alignment and route tracking. Add LiDAR if the job depends on precise 3D structure, not only defect finding.

Common mistakes and anti-patterns

1. Comparing headline range only

Range without target reflectivity, field of view, resolution, and weather context is mostly marketing. Ouster’s own table shows how range, precision, and FOV trade against one another across OS0, OS1, and OS1 Max.

2. Ignoring timestamp alignment

A camera frame and a radar return that describe different instants are not a fused truth. Time sync is not optional if you want clean fusion or replayable logs.

3. Treating IMU output as ground truth

IMUs are excellent at short-term motion prediction, but the gyro drift problem is intrinsic. Use them to carry the estimate between corrections, not to define position forever.

4. Assuming GNSS always works outdoors

GPS.gov and FAA guidance both treat interference, spoofing, and masking as operational issues. If your stack fails whenever the sky is bad, you do not have a robust stack.

5. Using ultrasound for too much

Ultrasonic sensors are cheap and useful, but they are not precise 3D perception tools. Wide beams, soft surfaces, and short range make them a special-purpose proximity sensor.

6. Over-trusting camera AI alone

Vision is superb at classification and context, but it is brittle under glare, darkness, and occlusion. If the task needs distance or motion safety, pair it with LiDAR, radar, or IMU.

Primary sources and verification targets

Related pages