To achieve Level 4 and above autonomous driving, vehicle localization must remain accurate and reliable under diverse real-world conditions, including complex traffic scenarios, environmental changes, and partial sensor failures. Conventional localization approaches primarily rely on geometric consistency among multi-sensor observations, which can produce physically or contextually implausible pose estimates when sensor reliability degrades or observations become ambiguous. This paper proposes a semantic localization framework that integrates ontology-based semantic reasoning directly into the localization process. The proposed approach reformulates localization as a context-aware constraint selection problem guided by semantic consistency among objects, places, and vehicle poses. By evaluating logical and contextual validity at the hypothesis generation stage, semantically invalid pose hypotheses are eliminated early, and only situation-appropriate semantic constraints are selectively applied during optimization. As a result, compared to the localization system without semantic rules, the proposed framework achieves an average reduction of approximately 35.6% in mean localization error and 47.0% in maximum localization error across both longitudinal and lateral directions. Specifically, the framework supports structured multi-sensor fusion by selectively using sensor information semantically relevant to the driving context. Through this semantics-driven hypothesis reduction, the system reduces computational complexity while enhancing localization robustness and accuracy, particularly under sensor degradation and dynamic environmental conditions. The design of the semantic reasoning structure is also adaptable to cooperative perception scenarios, such as V2V-based information sharing.

Eccentric payloads redistribute wheel loads and change the forces and moments required to maneuver heavy mobile robots. This study combines torque-based center-of-mass (CoM) estimation with global and local motion planning. During aligned, quasi-static creep, normalized wheel torques give a planar CoM estimate under calibrated rolling-resistance assumptions. An acceleration-based extension for CoM height remains unvalidated on hardware. The global planner adds a direction-selective turning penalty, while the DWB-style local planner uses an operational rollover-index constraint and a preview speed limit. Tests on a 2000 kg platform with a 500 kg payload recovered the reference quadrant at all four asymmetric placements using forward–reverse creep. Combined-CoM discrepancies were 20–75 mm at three placements and 265 mm at the front-left placement. In fixed-placement simulations with nominal offsets, local planning reduced the peak index by 40.3% on average, and the combined global and local planners by 44.7%, without violating the selected bound. A reduced-multibody replay retained a 39.7% mean peak reduction at the nominal contact stiffness but produced brief violations at three placements. With a recovery policy that keeps the constraint active during normal candidate selection, 40 randomized scenarios had no violations; with hardware-informed error scales, ten exceeded the bound. Estimation accuracy and physical validation remain limitations.

Regularization can improve generalization by constraining how a model uses its internal representation. In this paper, we study whether algebraic fixed-point constraints applied to the final LSTM hidden state during training can reorganize the recurrent latent space and improve held-out frame prediction. Rather than modifying the inference-time architecture, we introduce four training-time operators, GlobalHouseholder (reflection), GlobalGivens (rotation), Composition, and Lie Algebra, that bias the hidden state toward geometrically structured regions without changing the decoder pathway. Experiments across four datasets (indoor robot sequences, KITTI driving, Flying Shapes 2D, and Moving 3D Shapes) show that lightweight constraints consistently improve prediction on structured scenes, with GlobalGivens achieving up to $+1.04$ dB PSNR and $-11.3$% MAE over the unconstrained baseline on held-out Indoor sequences. The latent analysis reveals that the operators that generalize best are not those that compress the representation most aggressively but those that redistribute latent energy while preserving broad dimensional participation. Lie Algebra, despite collapsing activation variance by 81–96%, degrades under latent perturbation and does not match the lighter operators on structured datasets identifying over-constraint as a clear failure mode. These results suggest that geometric regularization of a recurrent bottleneck can act as a useful training-time prior without adding any inference overhead.