Conflict-Weighted Meta-Gradient Correction for Continual Learning in Multi-Sensor UAV Positioning

Abstract

Multi-sensor UAV positioning systems are increasingly deployed in sequences where new sensor modalities are integrated mid-deployment, requiring a model to expand its input space while retaining knowledge of previously observed environments. Existing continual learning methods assume a fixed-dimensional input and fail in this modality-incremental settings with earlier single-modality task memories produce stale gradients that actively oppose learning of newly introduced sensor features. We propose RAMGCL, a method that addresses this dual challenge of catastrophic forgetting and input-space expansion through three components operating within a fixed-size architecture. An asymmetric modality filter selectively suppresses stale gradients from zero-padded memories while retaining cross-modality signal that remains informative. A conflict-weighted per-task memory gradient continuously corrects the update direction by upweighting tasks whose gradients most conflict with the current update. A bi-level diagonal gradient preconditioner meta-learns a per-dimension scaling vector for the latent gradient slice, calibrating stability and plasticity without solving a quadratic programme. Episodic memories are populated using farthest-point sampling in position space to ensure spatially diverse coverage of each environment. Experiments on the Q-Drone UWB benchmark across five environments and two modality-incremental orderings show that RAMGCL achieves state-of-the-art performance in both average MAE and forgetting, and is the only method to achieve positive backward transfer in the IMU-to-UWB ordering.