RobotGuide
Sensor selection/The route debate

Selection · The route debate · About 30 min

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

'Should we add a LiDAR' is not a technical question; it is a trade-off between failure modes and cost. This page lays out both sides' physical principles, failure modes and cost curves, gives per-task criteria - and explains why the industry still argues.

About 30 min2026 market view

One-line conclusion

LiDAR measures distance directly, does not depend on lighting and has stable precision, but fears rain, fog and dust and is expensive; vision is cheap, information-dense and has semantics, but depth must be inferred and it fears darkness and backlight. Their failure modes barely overlap - so the 'fusion camp' is structurally right for safety-critical scenarios, while the 'vision-only camp' wins on the cost curve.

One sentence of physics decides each side's Achilles' heel

LiDAR: it emits its own light (laser), measures the time the light takes to fly out and return, and gets distance directly. So it works in the dark and its precision does not change with the scene - but rain and fog scatter the light, and mirrors and glass bend it the wrong way. Its Achilles' heel is the 'medium' and 'reflective properties', not lighting.

Vision (mono / stereo / RGB-D): passively receives ambient light. Depth comes either from the parallax between two viewpoints (the stereo part of stereo / RGB-D) or from a model's guess (mono + neural network). So harsh outdoor daylight, low-texture white walls and backlight can all break it - its Achilles' heel is 'light conditions' and 'scene texture'. RGB-D's structured-light / ToF variants emit their own IR fill light and work in dim conditions, but outdoor sunlight drowns the IR signal.

The two sides' Achilles' heels barely overlap: one fears weather but not darkness, the other fears darkness and strong light but not weather. That is the structural reason for fusion - not 'two is better than one', but 'when one goes blind the other probably has not'.

Spec comparison (same price band, 2026 terms)

DimensionLiDAR (thousand-RMB domestic)RGB-D cameraStereo
Depth accuracy±2–3 cm, stable across full range±1% (±5 mm at 0.5 m, from ±2 cm at 2 m)Depends on resolution and baseline; fails outright in low-texture regions
FOV2D 360° / 3D around 120°×25°About 70–90°×55°Same as camera FOV; wider baseline is more accurate
Frame rate / latency10–20 Hz30 fps (depth 15–30 fps)Compute-limited; matching algorithms eat CPU/GPU
Point / pixel density2D a few hundred points per revolution; 3D hundreds of thousands of points/sMillions of pixels per frame, the densestSame as above
Outdoor performanceGood (except rain and fog)Poor (sunlight drowns IR)Medium (auto exposure must be tuned)
Night performanceGoodStructured light / ToF usable, stereo poorPoor (underexposed)
Compute burdenLow (point cloud given directly)Medium (driver + alignment)High (stereo matching)
PriceDomestic 2D 300–800, 3D from the thousand-RMB level500–3000200–1500 + compute cost

Why the industry has argued for ten years

The cost-curve divergence. Mechanical LiDARs were once tens-of-thousands-of-dollars components; after 2019 domestic solid-state solutions pushed prices down to the thousand-RMB level and kept pushing, while cameras were already hundred-RMB items. 'Is it worth spending more for an independent ranging channel' - the answer changes every year with the price curve. That is the economic reason the debate lasts.

The two routes' real point of divergence. The vision-only camp (Tesla as the publicly known representative) argues: vision carries the densest information, comes with semantics, dilutes its cost with volume, and humans drive with vision alone - redundancy should be patched by algorithms and data scale. The fusion camp argues: safety-critical systems should not bet that 'algorithms are reliable under every lighting condition'; an independent ranging channel of a different physical principle is insurance, and after LiDAR's price drop the premium keeps getting cheaper. Both are right; the disagreement is over how much safety redundancy is worth.

The reality in 2026. Most passenger-car advanced driver assistance uses fusion (LiDAR has entered mainstream models); low-speed delivery robots, cleaning robots and humanoid head products are mostly fusion; on humanoids LiDAR and depth cameras coexist (one 3D / solid-state on the head + depth cameras on the hands). For individual developers the answer is simpler: can your budget afford a second independent ranging channel? If yes, add it - the value of non-overlapping failure modes far outweighs a pretty spec sheet.

Per-task criteria (answers for developers)

Two hard recommendations for beginners

For your first setup buy the cheapest combo that 'runs the full pipeline', not the one with the prettiest specs. A few-hundred-RMB combo of 2D LiDAR + IMU + camera can run mapping → localization → navigation → obstacle avoidance end to end; that builds ability more than any single high-end piece of hardware.

Put 'failure modes' on your test checklist: test once with lights off, once with curtains drawn, once facing a mirror, once with a mist of water. A sensor combo that survives all four tests is truly usable.

Deeper water on this site

What to do after getting a point cloud: /k/E2/ (point cloud processing).

Multi-sensor timestamp and coordinate-frame alignment: /k/B1/ (tf tree) - fusion's first lesson is not algorithms, it is a 20 ms timestamp.

The math of Kalman fusion: /math/gauss/ (inverse-variance weighting of two rulers).

← Previous: the sensor selection master guide  ·  Hands-on layer: point cloud processing

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