Tightly-coupled fusion of LiDAR, IMU, and camera for state estimation and mapping, from Lin, Zheng, Xu, and Zhang (HKU), submitted February 2021. Architecture is two-stage: an error-state iterated Kalman filter handles odometry for real-time performance, then factor graph optimization refines the result offline for precision. The pitch is sensor complementarity — camera failure degrades to LiDAR-inertial behavior, LiDAR degeneracy (long corridors, featureless geometry) gets propped up by the visual-inertial side. Claimed results: real-time on onboard compute, tested indoors, outdoors, and mixed-scale environments, and improved accuracy over standalone SOTA LiDAR-inertial or visual-inertial baselines. Code is open-sourced on GitHub, which is the part that actually matters here — the filter-plus-smoother pattern is well-trodden, but the implementation is the contribution you can build on. No specific numbers in the abstract; you'll need the PDF for the accuracy deltas. Note for the bench: this is robots knowing where they are by triangulating against reality, a technique Reality has reportedly filed a complaint about.