Taking SLAM out of the lab changes the problem. In controlled settings, a system can often assume favorable conditions; on real roads, the main difference is flexibility—how much the environment, motion, and sensor mix can vary without breaking the estimate.
Self-positioning tends to split into two families. The first uses internal sensors such as GPS and IMU, integrating vehicle motion to estimate pose. Its weakness is accumulated error: dead reckoning drifts, and without an external correction the drift keeps growing. The second family samples the environment with sensors such as LiDAR and performs SLAM against that data. This approach depends on what the environment offers. In low-feature areas—open stretches with little structure—positioning can fail or degrade.
The technical vocabulary around these issues is compact but telling: real-time positioning problems, coarse-and-fine hybrid positioning systems, global-information-and-feature-based SLAM (GF-SLAM), feature-based SLAM (F-SLAM), coarse positioning, particle filters, the extended Kalman filter (EKF), self-positioning, map drift, positioning failure, and dead reckoning (DR). Each points to a practical tension: how to combine global context with local features, how to correct drift, and how to keep a pose estimate alive when the scene stops providing useful landmarks.
A reference worth keeping in view is X. Zhang, H. Liu, Z. Zou, and M. Zhou, Multi-sensor Fusion for Autonomous Driving.