R1-2500067
discussion
Discussion on AI/ML-based positioning enhancement
From ZTE
ZTE's prior position on
9.1.2
at
RAN1#118bis
· AI-synthesized, paraphrased
verify sources →
Strongly supports maximizing reuse of existing 3GPP procedures and advocates for CIR with phase information over PDP for better positioning accuracy despite higher overhead, while also favoring sample-based measurements over path-based approaches.
Summary
ZTE presents 32 proposals and 3 observations regarding AI/ML-based positioning enhancements for NR Rel-19, focusing on model input definitions, phase information utility, and monitoring procedures. The document argues for reusing existing legacy signaling structures (such as timestamps and quality indicators) to minimize specification impact while supporting sample-based channel measurements. It specifically addresses the trade-offs between Channel Impulse Response (CIR) and Power Delay Profile (PDP) inputs and defines strategies for model training data association and performance monitoring.
Position
ZTE proposes reusing existing legacy signaling structures, such as 'measurementReferenceTime' and 'NR-TimeStamp', to minimize specification impact for AI/ML positioning timestamps and quality indicators. They argue against introducing a new 'associated ID' for training/inference consistency, preferring explicit provision of legacy assistance data instead. ZTE supports the inclusion of phase information (CIR) as model input, presenting technical evidence that CIR offers superior positioning accuracy compared to PDP with acceptable overhead increases. They oppose reporting the transmit offset from gNB to LMF in Case 3b and argue that power quality indicators are unnecessary for channel measurements. For model monitoring, ZTE supports LMF-side metric calculation (Option B) for Case 1 and proposes that label-free monitoring be handled by implementation transparent to the specification.
Key proposals
- Proposal 1 (Model Input): Supports candidate values for Nt' (9, 16, 24) and k (0…5) for sample-based measurements in Case 3b.
- Proposal 2 (Model Input): Opposes reporting the transmit offset from gNB to LMF in AI/ML positioning Case 3b, arguing it carries no channel information.
- Proposal 3 (Model Input): Proposes enhancing Case 2b measurements with Nt' values of estimated channel response in time domain, selected from Nt consecutive values with timing granularity T.
- Proposal 4 (Model Input): Argues there is no need to introduce power quality for channel measurement, relying solely on timing quality.
- Proposal 6 (Model Input): Proposes reusing TSubframeRxi (Option 1) or TUE-TX (Option 2) as reference time for UE channel measurement in Case 2b.
- Proposal 7 (Model Input): Proposes using phase information (CIR) for determining model input, citing better positioning performance compared to PDP despite slightly higher overhead.
- Proposal 8 (Model Input): Proposes reusing the existing RSRPP mapping table in TS 38.133 for AI/ML positioning power measurements.
- Proposal 9 (Model Output): Proposes that for Case 3a, the LOS/NLOS indicator provides the likelihood of a line-of-sight physical propagation path when reported with legacy timing measurements.
- Proposal 11 (Model Output): Proposes that an indicator identifying whether a measurement is based on an AI/ML model is not needed for reporting model output.
- Proposal 13 (Model Training): Proposes reusing the existing IE 'measurementReferenceTime' in TS 37.355 for the timestamp of location information in Case 1.
- Proposal 18 (Model Training): Proposes reusing 'horizontalUncertainty' for location labels and 'NR-TimingQuality' for timing labels to indicate label quality.
- Proposal 22 (Model Training): Proposes associating Part A (measurements) and Part B (labels) using timestamps and UE IDs, with association determined by the model training entity.
- Proposal 23 (Consistency): Proposes providing all legacy DL-TDOA assistance information explicitly to the UE and opposes introducing an 'associated ID' for AI/ML positioning.
- Proposal 26 (Model Monitoring): Supports Option B-1 and B-2 for label-based model monitoring in Case 1, where the LMF performs metric calculation using inference results sent by the UE.
- Proposal 32 (UE Capability): Proposes that UE capability for AI/ML positioning is reported per use case (Case 1, 2a, 2b).