Lidar SLAM vs camera (visual) SLAM
Should my robot map the world with a laser or a camera?
The short answer
Pick lidar when you need reliable geometry (indoor navigation, warehouses, anything that must not bump into things); pick a camera when weight, cost or power rule lidar out, or when you also need to recognise what you see. Serious robots fuse both.
Both build a map while tracking the robot inside it. Lidar measures distance directly and works in the dark; a camera is cheaper, lighter and sees texture and objects, but has to infer distance and fails in low light or on blank walls.
| Aspect | Lidar SLAM | Visual SLAM | Edge |
|---|---|---|---|
| What the sensor measures | Distance to the nearest surface along each laser beam — direct range | Brightness per pixel — depth has to be inferred from parallax, stereo or a learned model | Lidar |
| Lighting | Works in total darkness; dazzled by direct sun on some units | Needs light; struggles with glare, darkness and fast exposure changes | Lidar |
| Featureless spaces (long corridors, white walls) | Corridors are ambiguous for 2D lidar, but walls still give range | Few visual features to track; drift and tracking loss are common | Lidar |
| Glass and mirrors | Lasers pass through or reflect off glass — classic lidar failure | Also confused by reflections, but texture behind glass can still be tracked | Even |
| Dynamic scenes (people walking past) | Moving blobs corrupt scan matching unless filtered | Moving objects corrupt feature tracking unless masked | Even |
| Semantic information (what is that?) | A point cloud has shape but no labels | Images carry colour, text, object identity — recognise a door, a charger, a person | Visual |
| Cost | 2D units from ~$100 (robot-vacuum class) to several thousand for 3D | A usable camera from ~$20; stereo or RGB-D from ~$150 | Visual |
| Size, weight, power | Spinning 2D units ~150–300 g; 3D units heavier and power-hungry | Grams, milliwatts — fits on a drone or a toy | Visual |
| Compute | Scan matching is cheap; 2D SLAM runs on a Raspberry Pi | Feature extraction and bundle adjustment are heavier; real-time needs a decent CPU or GPU | Lidar |
| Loop closure (recognising a place you have been) | Possible from geometry, but similar rooms look alike to a laser | Strong: place recognition from appearance is what cameras are good at | Visual |
| Scale | Metric by construction — the map is in metres | Monocular SLAM has no absolute scale without an IMU, stereo or depth sensor | Lidar |
| Mature open-source stacks | SLAM Toolbox, Cartographer, GMapping (2D); LIO-SAM, FAST-LIO (3D) | ORB-SLAM3, RTAB-Map (RGB-D), VINS-Fusion, OpenVINS | Even |
Choose Lidar SLAM when…
- Indoor mobile robots that must navigate reliably: warehouses, hospitals, offices.
- Anything that operates at night or in windowless rooms.
- You want a metric 2D map you can hand to Nav2 today.
- Compute is small (a Pi or a Jetson Nano) and the budget allows a ~$100+ 2D lidar.
Choose Visual SLAM when…
- Weight or power is critical: drones, small wheeled robots, wearables.
- You also need to recognise objects, read signs or find a specific thing.
- Budget is tens of dollars, not hundreds.
- You can add an IMU (for scale and robustness) — visual-inertial SLAM is far stronger than a bare camera.
What most real robots do
Robot vacuums show the split at consumer scale: Roborock, Dreame and most Xiaomi models carry a small spinning lidar in the turret; iRobot's Roomba j-series uses a front camera instead. Above that price point the argument mostly disappears — Spot, warehouse AMRs and self-driving cars run lidar and cameras and an IMU through a sensor-fusion stack, because each sensor covers the other's blind spot.
Plain English. Free. Unsubscribe in one click.