DRL-Based Beam Positioning for LEO Satellite Constellations with Weighted Least Squares

TL;DR

This paper presents a lightweight deep reinforcement learning framework combined with weighted least squares for accurate, resource-efficient satellite positioning in LEO constellations, achieving sub-meter accuracy with low computational cost.

Contribution

It introduces a novel hybrid DRL and WLS approach for satellite positioning that balances accuracy and computational efficiency in resource-constrained environments.

Findings

Achieves 0.395m RMSE in 2-D satellite positioning.

Supports practical deployment with low computational overhead.

Provides physics-consistent localization and joint clock bias estimation.

Abstract

This paper investigates a lightweight deep reinforcement learning (DRL)-assisted weighting framework for CSI-free multi-satellite positioning in LEO constellations, where each visible satellite provides one serving beam (one pilot response) per epoch. A discrete-action Deep Q-Network (DQN) learns satellite weights directly from received pilot measurements and geometric features, while an augmented weighted least squares (WLS) estimator provides physics-consistent localization and jointly estimates the receiver clock bias. The proposed hybrid design targets an accuracy-runtime trade-off rather than absolute supervised optimality. In a representative 2-D setting with 10 visible satellites, the proposed approach achieves sub-meter accuracy (0.395m RMSE) with low computational overhead, supporting practical deployment for resource-constrained LEO payloads.