Semantic Communication for Closed-loop Sensing-Communication-Control Systems

Abstract

This thesis investigates channel state information (CSI) as a sensing input for position estimation and simulated closed-loop control, motivated by the Level-2 interpretation of semantic communication. It evaluates task-relevant state extraction rather than a modeled semantic transmission link. A simulation pipeline combines Sionna RT channel generation in a street-canyon scene, a ResNet-based estimator, Kalman filtering, and linear quadratic regulator (LQR) control. Complex CSI is represented by real and imaginary components across antennas and subcarriers. The neural model predicts two-dimensional position and coordinate-wise variance. Position predictions supply measurements to the filter, while the controller acts on the filtered state. The learned covariance is evaluated separately; the main closed-loop experiments use fixed diagonal covariance. On a random held-out test split of 2000 samples within the simulated scene, the estimator achieves a mean Euclidean localization error of 0.857 m, an RMSE of 1.204 m, and a median error of 0.673 m. Joint antenna–subcarrier masking yields an accuracy plateau: retaining 56.25% of the full CSI coefficients gives the lowest observed mean error, 0.626 m. This does not establish bitrate or energy savings. Predicted covariance is weakly correlated with localization errors and fails the reported consistency check, limiting its direct use in Kalman filtering. The fixed-covariance sweep demonstrates tracking sensitivity to measurement confidence. In one paired seed-0 rollout per setting, a variance of 1 m2 per coordinate yields the lowest true-state LQR cost, 750.27, and tracking RMSE, 0.683 m, among the tested settings. These findings characterize within-scene localization, uncertainty limitations, and covariance-sensitive tracking. Ideal interfaces, one scene and trajectory, and missing controller controls restrict conclusions about communication benefits, generalization, and the contribution of individual pipeline components. Calibrated uncertainty and broader evaluation are needed before deployment.