R1-2409500
discussion
Discussion on specification support for positioning accuracy enhancement
From CMCC
Summary
CMCC analyzes specification impacts for AI/ML-based positioning in NR, presenting 17 proposals and 12 observations across methodology, measurement definitions, data collection, and model monitoring. The document argues for sample-based measurements over path-based ones to reduce processing ambiguity and emphasizes the need for offline data collection with minimal spec impact.
Position
CMCC slightly prefers sample-based measurements for AI/ML positioning to avoid the potential errors introduced by intermediate path-based processing at the UE. They propose that the entity deriving the AI model should provide the recommended number of samples (Nt') and that selection rules, such as strongest power samples or thresholds, should be specified to reduce ambiguity. For data collection, CMCC argues that offline training with marginal specification impact is feasible, but questions the feasibility of obtaining ground-truth labels via PRUs due to large dataset sizes. They propose reinterpreting the existing 'Timing Measurement Quality' IE to report performance metrics for gNB-side models and suggest that legacy LOS/NLOS reporting formats be reused. Regarding consistency between training and inference, CMCC deprioritizes options 1 and 2, preferring further discussion on options 3 and 4, and emphasizes that the granularity of validity areas needs further study. For second-priority use cases, they require that reported timing information for Case 2a be based on legacy reporting formats to minimize LMF modifications.
Key proposals
- Proposal 1 (Methodology): Sample-based measurements for AI/ML positioning is slightly preferred over path-based measurements to avoid intermediate processing errors.
- Proposal 2 (Methodology): For selecting Nt' samples by UE, assuming that these Nt' samples are the strongest power samples in the sampling window or introducing a threshold could be considered.
- Proposal 3 (Methodology): For selecting Nt' samples, the entities that derive AI model could provide the recommended number of samples.
- Proposal 4 (Considerations): For path-based measurement reporting, current DL PRS-RSRPP and UL-SRS-RSRPP can be used as a starting point.
- Proposal 5 (Considerations): For AI/ML based positioning, it needs more discussion on the feasibility of obtaining the ground-truth label via PRUs, in which case the training dataset size is large.
- Proposal 6 (Considerations): For AI/ML based positioning, more discussion is needed for the comparison between CIR and PDP as model inputs.
- Proposal 7 (Considerations): For AI/ML based positioning, additional configurations or limitations can be supported to improve the efficiency of the measurement data reporting for positioning.
- Proposal 8 (Considerations): For UE generates ground truth label based on non-NR and/or NR RAT-dependent positioning methods, the reliability or the positioning accuracy should also be reported.
- Proposal 9 (Considerations): For AI/ML based positioning, whether the reported measurement is AI based could have an indication.
- Proposal 10 (Considerations): For AI/ML assisted positioning, existing IE “Timing Measurement Quality” can be reinterpreted for reporting performance metric for gNB side model.
- Proposal 11 (Considerations): For AI/ML assisted positioning, legacy LOS/NLOS information reporting could be used.
- Proposal 12 (Considerations): For performance monitoring is based on the ground-truth labels, further study method to obtain ground-truth label.
- Proposal 13 (Considerations): For AI/ML based positioning, the relationship between model monitoring and positioning integrity can be considered.
- Proposal 14 (Considerations): For AI/ML based positioning, further discuss option 3 and 4. Deprioritize option 1 and 2 regarding consistency between training and inference.
- Proposal 16 (Additional spec impact): For AI/ML assisted positioning case 2a, the reported timing information should be based on the legacy reporting format.