Shared-Encoder Reinforcement Learning for Multitask LiDAR–Inertial SLAM

Zhaotian Deng, Ji Li, Qingjiao Meng, Weixin Zhang, Jicheng Yao, Chuchu Zhu

Keywords: SLAM, LiDAR-inertial odometry, Reinforcement learning

TL;DR: A shared-encoder multitask reinforcement learning controller jointly performs IMU denoising, adaptive keyframe selection, and loop closure triggering to enhance LiDAR–inertial SLAM accuracy and efficiency.

Abstract:

LiDAR–IMU SLAM in long-term operation is still
exposed to IMU drift, redundant keyframe insertion and unsafe
loop-closure attempts, which jointly degrade pose accuracy and
increase computational load. We integrate a multi-task reinforce-
ment learning controller into the SLAM pipeline to address these
coupled failure modes. The controller observes inertial statistics
over a sliding window together with estimator load indicators
and lightweight loop-closure cues. A shared encoder produces
a compact latent state that drives three task heads: IMU bias
denoising, adaptive keyframe selection and selective loop-closure
triggering. The heads output bias correction for pre-integration,
a keep or skip decision for each keyframe candidate and a
trigger decision for loop verification, which are injected directly
into the factor-graph back-end. This design improves accuracy,
stability and real-time efficiency without altering the underlying
optimization structure. On KITTI, compared with the baseline,
our method reduces translational error by 6% and rotational
error by 0.7%.

Submission Number: 63