Past, Present, and Future of Simultaneous Localization And Mapping: Towards the Robust-Perception Age

Overview of SOTA SLAM methods and pipelines (SLAM: Simultaneous Localization and Mapping - estimate robot state and environment map jointly) from 2016. Classical age SLAM (1986 to 2004) involved extended Kalman filters (EKFs), Rao-Backwellised Particle filters (RBPFs), and maximum likelihood estimation (MLE). Algorithmic analysis age (2004 to 2015) involved convergence, consistency, etc. SLAM relates to: Odometry and loop closure, sensor fusion under challenging conditions, visual inertial navigation (VINs) systems, visual place recognition. Requirement based on type of robot, operation environment, and performance requirements. Architecture: front-end takes sensor data and does feature extraction and association; back-end does maximum a posteriori (MAP) estimation through an optimization method (over a factor graph) by taking negative log (assuming Gaussian/normal noise) - non-linear least squares problem (sparse and can be solved using solvers like GTSAM, G2O, Ceres, iSAM, SLAM++, etc.). Like bundle adjustment (BA) in SfM but can include models outside projective geometry (sensor and motion models). Front end takes care of feature/landmark extraction in the map, loop closure detections, data associations, and can also provide initialization for landmarks (by triangulation). Robustness: incorrect data association can be addressed through RANSAC, but back-ends are still not failure safe (and failure aware); hardware failure (retaining operation under degraded sensor), non-rigid maps, and automatic parameter tuning still under research. Long term SLAM (unbounded factor graph through exploration) does (graph) sparsification or computation distribution over multiple robots; map representations for perpetual learning, forgetting and remembering map portions, and robust distributed mapping are under research. Metric map models could be landmark/feature based (keypoint centric mapping), locally dense (point cloud, polygon meshes), boundary based (surface representations like SDFs), or even 3D object representations. Expressive, optimal, and adaptive representations still a challenge. Semantic information helps SLAM mapping and optimization. Semantic metric fusion, adaptive mapping, and knowledge of environment in semantic (and metric) map updates still under research. Active SLAM controls robot’s motion to minimize uncertainty of map representation and localization; exploration for information gain (get vantage points and explore one with highest utility); forecasting considering all actions seems intractable. TOF sensors (range and LiDAR), and event cameras are new sensor modalities. Deep learning is mostly being used for perception (front end). SLAM review from ETHz, MIT (Luca Carlone), University of Adelaide, University of Zurich.

Links: Website, GitHub