Laser SLAM for Service Robots: Comparing Approaches

A laser SLAM service robot knows where it is by matching rangefinder data against a map it has built, and the number of scan planes it uses changes almost everything downstream, from the compute budget to the kind of obstacle it can avoid. The three approaches seen in the field are a single plane two dimensional scanner, a multi channel three dimensional scanner, and a fusion of lidar with inertial, wheel and camera data. This article compares them on the points that decide a deployment: what each one sees, what each one misses, and which buildings justify the extra cost.

Two Dimensional Scan Matching

A single plane scanner mounted at knee height is the workhorse of indoor service robots. It produces one slice of the world at a high update rate, and because the data is thin, scan matching plus a particle filter runs comfortably on embedded hardware. The resulting map is a two dimensional occupancy grid that a facilities team can read and edit, and localization is stable in corridors, hospital wards and retail aisles where walls repeat but remain individually distinguishable. The weakness is vertical blindness. Glass and mirrored surfaces return weak or missing echoes, pallets, cables, door sills and low furniture pass underneath the scan plane, and counters, tables and open drawers pass above it. Ramps and slopes also distort the picture, because the scan plane is no longer parallel to the floor.

Three Dimensional Lidar and Its Real Cost

A multi channel lidar adds height and removes most of that blindness. It notices a pallet lying flat, a forklift tine, an overhanging shelf and the slope of a dock ramp, which is why outdoor delivery platforms and warehouse vehicles working in mixed traffic tend to rely on it. The costs are equally real. Point clouds arrive at a far higher rate, the sensor head is heavier and more expensive, and both matter on a robot whose payload and battery are limited. Registration between successive scans is also more delicate, because in a long featureless corridor or an open plaza the geometric constraints are weak in the direction of travel and drift accumulates until a loop closure corrects it. A common compromise is to keep a three dimensional sensor for perception and obstacle avoidance while the planner runs on a two dimensional costmap projected from it, which keeps navigation simple and predictable.

Fusion with Odometry and Inertial Data

No lidar works alone in a well designed platform. Wheel encoders give smooth short term motion estimates that bridge the interval between scans, an inertial unit captures rotation that scan matching resolves poorly, and depth cameras add semantics such as a person, a cart or a chair. The coupling can be loose, where each sensor contributes a pose that a filter merges, or tight, where the measurements enter the estimator directly. Tight coupling copes better with fast rotation and sparse geometry, at the price of harder calibration and harder fault diagnosis when something drifts. Time synchronization is where many integrations quietly fail, because a stamp offset of a few milliseconds between the lidar and the inertial unit smears the map as soon as the robot turns. Extrinsic calibration also has to be rechecked after any mechanical repair, since a bracket bent by a fraction of a degree leaves a permanent offset in every map the robot builds.

Matching the Approach to the Environment

The building usually decides more than the algorithm does. Floors that are flat, textured and moderately busy are well served by a single plane. Sites with glass walls, outdoor aprons, loading docks and pallet handling need height, or they need the route designed so that the blind zone never faces a hazard. Crowded public spaces call for extra sensing and for a planner that treats people as moving obstacles rather than as noise to be filtered out. Weight, compute and battery budget then set the ceiling on what can be carried. One boundary deserves stating plainly, which is that the navigation sensor is not a safety device. Personnel protection must come from scanners and control functions certified for that purpose, and it has to stay effective even when the localization layer is confused or the map is stale.

Before committing to a configuration, walk the planned route with a tape and a laser level and note every reflective surface, low obstacle and slope, then decide whether one plane can survive it. Ask for a demonstration on your own floor rather than in a cleared test area where the map has no surprises. Put map maintenance and recalibration intervals into the acceptance criteria, because a stale map is one of the most common reasons a platform that worked for a year suddenly starts stopping in the wrong places.