Robust drift-free inertial initialization under high-dynamics motion

Develop robust, drift-free inertial initialization for high-dynamics motion in monocular-inertial ORB-SLAM3, to reduce sensitivity to under-converged initialization and prevent initialization errors from dominating trajectory accuracy on difficult sequences.

Background

The paper evaluates an accuracy-preserving GPU implementation of ORB-SLAM3 on an NVIDIA Jetson Orin Nano. In monocular-inertial operation, the system achieves real-time performance, but the authors report that accuracy on harder sequences is sensitive to IMU initialization. A single under-converged initialization can substantially increase the absolute trajectory error, and reporting multiple runs only mitigates rather than eliminates this variability.

The unresolved issue concerns the ORB-SLAM3 inertial-estimation pipeline itself rather than the GPU front end. Under aggressive motion, run-to-run variation and initialization failures remain sufficiently large that robust initialization capable of maintaining drift-free performance is identified as an open problem.

References

The monocular-inertial mode is real-time, but its accuracy is sensitive to IMU initialization on the harder sequences, where a single under-converged initialization can dominate the ATE. Multi-run reporting mitigates but does not remove this, and robust, drift-free inertial initialization for high-dynamics motion remains an open problem.

— Jetson-ORB-SLAM3: Accuracy-Preserving GPU Implementation for Edge Computing Devices  (2608.17874 - Roy et al., 18 Aug 2026) in Section 5, Observations, item (i), Section 5.1?