Perception and Sensors
LiDAR vs Stereo vs Time-of-Flight for Robots
Four 3D sensing technologies compared on range, accuracy, surface tolerance and cost, with the surface types that defeat each one.
Choose by surface, then range, then cost. Structured light gives the best short-range accuracy, typically 0.05 to 0.5 mm at 0.5 m, and fails on shiny and transparent parts. Time-of-flight covers 0.3 to 10 m at 1 to 10 mm and struggles at edges. LiDAR reaches 100 m and beyond with millimetre precision on a sparse point set. Stereo is cheapest and needs texture that many industrial parts do not have.
The four technologies
| Technology | Range | Accuracy | Rate | Fails on |
|---|---|---|---|---|
| Structured light | 0.2 to 2 m | 0.05 to 0.5 mm | 1 to 30 Hz | Shiny, transparent, dark, bright ambient light |
| Laser triangulation | 0.05 to 1 m | 0.005 to 0.1 mm | line scan | Needs relative motion, occlusion at steep angles |
| Time-of-flight camera | 0.3 to 10 m | 1 to 10 mm | 15 to 60 Hz | Edges, corners, multipath, dark surfaces |
| Passive stereo | 0.3 to 20 m | 1 mm to 50 mm | 15 to 90 Hz | Featureless surfaces, repetitive patterns |
| Active stereo | 0.3 to 10 m | 0.5 to 20 mm | 15 to 90 Hz | Bright sunlight, very shiny surfaces |
| 2D LiDAR | 5 to 40 m | 10 to 30 mm | 10 to 50 Hz | One plane only, glass, dark absorbers |
| 3D LiDAR | 20 to 200 m | 10 to 30 mm | 10 to 20 Hz | Sparse at distance, glass, rain and dust |
Note that accuracy and range trade against each other consistently. Nothing on the list gives 0.05 mm at 10 m, and any specification demanding it should be revisited before a sensor search begins.
Matching sensor to task
| Task | Accuracy needed | Sensor |
|---|---|---|
| Bin picking, mixed metal parts | 0.5 to 2 mm | Structured light or active stereo |
| Weld seam tracking | 0.1 to 0.3 mm | Laser triangulation |
| Palletising, case detection | 5 to 20 mm | Time-of-flight or stereo |
| Mobile robot obstacle detection | 20 to 50 mm | 2D safety LiDAR plus depth camera |
| Outdoor navigation | 30 to 100 mm | 3D LiDAR |
| Dimensional inspection | 0.01 to 0.1 mm | Laser triangulation or structured light |
| Human detection in a cell | safety rated | Certified safety scanner, not a depth camera |
The last row is a legal point, not a technical one. A depth camera can detect a person and is not a protective device. Only a certified safety sensor with a rated performance level may be relied on for a protective function, whatever the point cloud looks like.
What each category costs
| Sensor | Price band | Integration effort | Hidden cost |
|---|---|---|---|
| Passive stereo module | 200 to 1,500 | low | Texture projector often needed anyway |
| Active stereo camera | 400 to 3,000 | low | Ambient light shielding |
| Time-of-flight camera | 500 to 4,000 | low | Multipath correction and filtering |
| Industrial structured light | 4,000 to 20,000 | medium | Calibration fixture and lighting control |
| Laser triangulation profiler | 3,000 to 25,000 | high | Motion axis or robot-mounted scan path |
| 2D safety scanner | 1,500 to 5,000 | medium | Field calculation and validation |
| 3D LiDAR | 1,000 to 15,000 | medium | Point cloud processing compute |
The hidden cost column is where projects overrun. A 4,000 unit structured light head with a 6,000 lighting enclosure and two weeks of calibration work is a different purchase from the one in the quotation.
The mistakes that cost projects
- Specifying accuracy without range. Every figure in a datasheet is stated at a distance, and it degrades with the square of it on triangulation-based systems.
- Ignoring ambient light. A structured light sensor that works on a bench fails under a skylight. Sunlight near a loading door is the classic case.
- Testing with clean parts. Oil, coolant, chips and release agent all change how a surface returns light.
- Confusing precision with accuracy. A sensor can be repeatable and biased, which shows up as a consistent offset that calibration removes and noise that it does not.
- Forgetting the mount. A sensor on a vibrating frame delivers noise no filter can remove, and depth quality is often a mechanical problem.
Frequently asked questions
Which 3D sensor is most accurate?
Laser triangulation, at roughly 0.005 mm to 0.1 mm, followed by structured light at 0.05 mm to 0.5 mm. Both work only at short range, and accuracy degrades sharply with distance.
Why do shiny parts break depth sensors?
Because specular surfaces reflect the projected pattern away from the receiver instead of scattering it back. Transparent parts let it pass and matte black absorbs it. All three produce missing or noisy returns that software cannot recover.
Can I use a depth camera as a safety sensor?
No. Protective functions require a certified safety device with a rated performance level. A depth camera may detect a person, but it carries no safety claim and cannot be relied on for a protective stop.
Stereo or time-of-flight for a mobile robot?
Usually both, alongside a 2D safety scanner. Time-of-flight gives dense short-range depth for low obstacles, stereo covers wider range and works in daylight, and neither replaces the certified scanner.
How much accuracy does bin picking need?
Typically 0.5 mm to 2 mm on the pick pose, which structured light and active stereo both reach at working distance. The harder requirement is usually reliable segmentation of touching parts rather than raw depth accuracy.
Sources
- ISO 9283, manipulating industrial robots, performance criteria and test methodsAccuracy terminology used for sensor and robot specifications
- ISO 10218-1:2025, Robotics, Safety requirements, Part 1International Organization for Standardization, requirements for protective devices
- Robotics at NISTNational Institute of Standards and Technology, robot performance measurement and test methods