RobotGuide
Sensor selection

Selection · Selection logic · About 25 min

Sensor selection logic: ask what to measure first, then ask how much it costs

LiDAR, depth cameras, stereo, IMU, torque, tactile - the first step of selection is never comparing spec sheets, it is answering 'in this task, which information keeps the robot alive and which is a nice-to-have.' This page gives a five-step decision method and an all-category comparison table.

About 25 min2026 market view

One-line conclusion

Selection = physical quantity → range & precision → frame rate & latency → failure modes → budget & fusion. Only after these five steps does a spec sheet make sense. Skipping failure modes and going straight to comparing specs is the number-one way beginners die in selection.

The five-step method: each step eliminates a batch of options

Step 1 · What physical quantity to measure. Obstacle avoidance needs 'distance' (ranging types: LiDAR / depth camera / ultrasonic / stereo), localization needs 'pose change' (IMU / wheel odometry / laser matching), grasping needs 'contact and force' (torque / tactile arrays). First break the task into a list of information requirements; do not start with 'buy a lidar'.

Step 2 · Is range and precision enough. Indoor obstacle avoidance needs 0.5–10 m, outdoor autonomous vehicles need 200 m; grasp positioning needs ±1 mm, navigation ±5 cm is enough. Watch vendor terminology: LiDAR specs quote 'ranging accuracy' (single point), depth cameras usually quote 'Z-axis accuracy' (degrades with distance squared) - the two 'accuracy' figures on the same sheet are not the same thing.

Step 3 · Frame rate and latency. High-speed obstacle avoidance needs ≥20 Hz and end-to-end <50 ms; mapping is fine at 5 Hz. IMUs need 200–1000 Hz (they 'bridge time' between two slow sensors); cameras run 30 fps but exposure + transfer + algorithm pipeline often adds 80–150 ms - spec sheets give frame rate but not pipeline latency, the number-two way beginners get burned.

Step 4 · Failure modes (the most critical step). Every sensor has weather and scenes that 'kill' it: LiDAR fears rain, fog, dust and mirrors; cameras fear darkness, backlight and low texture; ultrasonic fears multipath and soft materials; IMUs fear temperature drift. Ask yourself one question: when this failure happens, will the robot hit a person? If yes - you must pair it with another sensor of a different physical principle for cross-validation. That is the entire rationale of 'redundant fusion'.

Step 5 · Budget and fusion plan. Spend the remaining money on the 'safety-critical chain', not on resolution. A cheapest workable combo (2026 terms): 2D LiDAR (300–800 RMB) + IMU (tens of RMB) + camera (100–300 RMB) is enough to run the full indoor mapping-and-navigation stack.

The order of the five steps cannot be swapped: comparing price before thinking about failure modes is the most common path to a crash - a cheap sensor failing exactly in the scene you need most equals having installed nothing.

All-category comparison table (common 2026 terms)

SensorWhat it measuresTypical specsAchilles' heel (failure mode)PriceTypical tasks
Mechanical spinning LiDARDistance point cloud10/16/32 lines, 10–20 Hz, ±2–3 cmRain, fog, dust, specular reflection, glassDomestic 2D 300–800 RMB; 3D 1500–tens of thousandsSLAM mapping, navigation and obstacle avoidance
Solid-state LiDARDistance point cloudNo rotating parts, FOV around 120°×25°Near-field blind zone, vibration misalignmentThousands to tens of thousands of RMB, price falling fastAutonomous vehicles, humanoid robot heads
RGB-D depth camera (structured light / ToF)Color + depth images30 fps, 0.3–5 m, Z accuracy degrades with distanceStrong sunlight (IR drowned out), black and mirror objects500–3000 RMBRobotic-arm grasping, indoor obstacle avoidance
Stereo cameraDisparity → depthAccuracy set by resolution; low texture breaks itLow texture, drastic lighting change, calibration drift200–1500 RMB (compute not included)Outdoor navigation, low-cost solutions
Monocular camera + modelSemantic / pseudo-depthNo true scale (single frame)Scale drift, model dependence100–300 RMBSemantic perception, tracking
Event cameraAsynchronous brightness changesµs-level latency, high dynamic rangeNo output in static scenes, small ecosystemThousands to tens of thousandsHigh-speed tracking, shake suppression (research-oriented)
IMU (6-axis / 9-axis)Angular velocity + acceleration200–4000 Hz, bias temperature driftThermal drift accumulation, magnetometer easily disturbedTens to thousands of RMBAttitude estimation, bridging time for slow sensors
Wheel encoders / odometryDisplacement incrementsSlip and it is wrongSlip, slip, slipTens to hundreds of RMBChassis odometry
Joint torque sensorTorque±tens of N·m, 1 kHzTemperature drift, overload damageThousands of RMB per joint (often built into integrated joints)Force control, collision detection
Tactile array (piezoresistive / capacitive / visuotactile)Contact distributionmm-level spatial resolutionWear, complex calibrationHundreds to thousands of RMBGrasp slip detection, fine manipulation

Three most common tasks, three common answers

Three high-frequency beginner traps

Comparing accuracy from vendor spec sheets alone: ranging accuracy and Z-axis accuracy are different metrics, and a depth camera's '±2 mm' usually only holds within 0.5 m.

Ignoring pipeline latency: camera 30 fps ≠ results in 33 ms. The whole chain of exposure → transfer → algorithms often measures 80–150 ms in practice - not nearly enough for high-speed obstacle avoidance.

No cross-validation of sensor failures: perception from a single physical principle is 'fully blind' inside its failure mode - no algorithm can patch that; it is a missing piece at the architecture level.

Related on this site

How sensor data fits together with coordinate frames: /k/B1/ (tf tree) - what a 20 ms timestamp gap between two sensors looks like.

After you get a LiDAR point cloud: /k/E2/ (point cloud processing).

Model sensor noise in simulation before buying hardware: /k/G1/ (Gazebo and production-grade simulation).

Read next → LiDAR vs vision: the selection take on the industry's ten-year debate

NextLiDAR vs vision: the selection take on a ten-year debate