R1-2500060 discussion

AI/ML for Positioning Accuracy Enhancement

From Ericsson
Status: not treated
WI: NR_AIML_air
Agenda: 9.1.2
Release: Rel-19
Source: 3gpp.org ↗
Ericsson's prior position on 9.1.2 at RAN1#119 · AI-synthesized, paraphrased
verify sources →
Presents a technical case against path-based measurements and CIR/phase inputs due to signaling overhead and alignment difficulties, initially favoring sample-based measurements with total-power PDP inputs. Later documents advocate supporting both sample-based and path-based measurements as a compromise to avoid blocking progress, while continuing to push against CIR support. Proposes using associated IDs or explicit assistance data from the LMF to ensure consistency between training and inference, and supports label-free self-monitoring by the model inference entity.

Summary

Ericsson presents a comprehensive contribution on AI/ML for NR positioning accuracy enhancement, focusing on integrating AI/ML methods with existing protocols, defining model inputs/outputs, and establishing training data collection and monitoring frameworks. The document contains 66 proposals and 38 observations across various sections, addressing protocol integration, signaling enhancements for model inputs, model output reporting, training data metadata, and lifecycle management.

Position

Ericsson proposes integrating AI/ML positioning into existing DL-TDOA and UL-TDOA/multi-RTT frameworks rather than defining new procedures, assigning procedural decisions to RAN2 and RAN3. They require the use of total-power PDP inputs summed over all receive antenna ports, explicitly opposing multi-port PDP due to doubled signaling size and marginal performance gains. Ericsson presents a technical case against supporting phase information or CIR inputs for model inference, citing random initial phase issues, high signaling overhead, and difficulty in aligning training/inference phases. They propose decoupling the LOS/NLOS indicator (reflecting physical LOS) from timing information (reflecting virtual LOS) in Rel-19 Case 3a reports. For training data, they require mandatory time stamps for unambiguous Part A/B pairing and support only UE location coordinates as ground truth labels. They establish self-monitoring as the baseline for model performance, with the inference entity responsible for metric calculation, while LMF handles functionality-level management. Finally, they propose using associated IDs to verify consistency of network-side conditions between training and inference.

Key proposals

Your notes

Private to your account