Localization: Where Am I?

Before a robot can navigate anywhere, it needs to answer a deceptively simple question: Where am I right now? This problem is called localization — determining the robot's own position and orientation within a known space or coordinate system. It sounds trivial until you realize that sensors are noisy, the world is dynamic, and even tiny errors in position estimates can compound into large navigational mistakes over time.

Odometry: Counting Wheel Turns

The simplest localization strategy is odometry — tracking how far the robot has moved by counting wheel rotations. If a wheel has a circumference of 50 centimeters and the encoder records 3 full rotations, the robot has traveled 150 centimeters in the direction the wheel was pointing. By combining the rotation counts from left and right wheels, the robot can also estimate how much it has turned. Odometry is inexpensive and requires no external landmarks, but it accumulates error relentlessly. Wheels slip on slick floors. Carpet fibers add slight, unpredictable drag. Manufacturing differences between left and right wheels create tiny biases. After a hundred meters, these small errors stack up into position estimates that can be several meters off. This growing uncertainty is called odometric drift.

Landmarks: Anchoring Position Estimates

One way to fight drift is to use landmarks — recognizable features in the environment whose positions are already known. When the robot's sensors detect a landmark, it can compare the observed position of the landmark with where that landmark should be on its map, and use the difference to correct its estimated location. Landmarks can be natural (a distinctive wall corner, a unique doorway) or artificial (a QR code sticker placed deliberately on the wall). Many warehouse robots use a floor covered in special markers — barcodes printed on tiles — that cameras read continuously. Every time a marker is detected, the robot's position estimate is snapped back to the known position of that marker, eliminating drift almost entirely.

Probabilistic Localization

A powerful modern approach to localization treats the robot's position not as a single definite point but as a probability distribution — a cloud of possibilities, each with an associated likelihood. At the start, before any sensor readings, the cloud might be spread uniformly across the entire map, reflecting total uncertainty. As sensors provide evidence, the cloud narrows: positions that are inconsistent with the sensor observations get lower probability, and positions that match get higher probability. One elegant implementation is called a particle filter. The robot maintains a population of hypotheses — thousands of virtual copies of itself placed at different positions. Each particle represents one guess about where the robot is. When sensor data arrives, particles whose hypothetical positions are consistent with the data are kept and replicated; inconsistent particles are discarded. Over time, the surviving particles converge on a small cloud around the true position.

GPS and Its Limits

Outdoor robots can use GPS (Global Positioning System) — a network of satellites that broadcasts precise timing signals from which a receiver can triangulate its position on Earth's surface. Standard GPS is accurate to about 3 to 5 meters, which is fine for car navigation but dangerously coarse for a robot that needs centimeter precision. Differential GPS and Real-Time Kinematic (RTK) GPS use ground-based correction stations to achieve centimeter-level accuracy outdoors. Indoors, GPS signals are blocked by building materials. In dense cities, signals reflect off skyscrapers — a problem called multipath error — producing readings that can be tens of meters off the true position. For these environments, robots rely on lidar, cameras, and the probabilistic methods described above.