Answers are grounded in this paper and cite their sources.
Multi-sensor SLAM fuses sensors that sample at different rates with unknown clock offsets, so the pose needed to correct a scan or associate a descriptor rarely coincides with a sampling instant. Existing methods interpolate discrete poses or evaluate a continuous-time trajectory inside one estimator, both taking the offset as known, so the offset error enters the estimate. We present a system built on a single IMU-driven cumulative B-spline trajectory. That trajectory is queried by preprocessing, the sliding-window estimator and the loop-closure module alike at the instant required for each measurement, and the inertial, LiDAR and visual terms all act on its control points. Both offsets are estimated online in the same optimization; the LiDAR–IMU offset is validated on the NCD by injecting known offsets. Because every descriptor carries its observation instant, a loop candidate is associated with LiDAR features by timestamp, which removes the two stages that only recover the relative pose of candidate frames and shortens the remaining registration. On NTU-VIRAL, our method attains a mean ATE of 0.151 m at 33.03 ms of LiDAR processing per frame, and the lowest mean ATE of 4.33 m on KITTI, with a 37% lower per-candidate loop-closure cost than LVI-SAM.
No takes yet. Share an insight, caveat, or question.
Ma et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: