R1-2410258
discussion
Discussion on specification support for positioning accuracy enhancement
From ETRI
Summary
ETRI presents a comprehensive technical contribution on AI/ML for NR air interface positioning accuracy enhancement, covering model input/output formats, training data collection, and performance monitoring across different positioning cases. The document contains 21 detailed proposals addressing various aspects of AI/ML positioning implementation in 3GPP specifications.
Position
ETRI strongly advocates for vendor flexibility in AI/ML positioning implementation, pushing for UE vendors and gNB manufacturers to determine their own model input formats rather than enforcing standardized formats. They support Option B for measurement window timing (earliest detectable sample) over other alternatives, and emphasize the need for separate handling of timing information and LOS/NLOS indicators in AI/ML-assisted positioning. ETRI is particularly focused on practical implementation considerations in NLoS environments where traditional positioning methods fail.
Key proposals
- Proposal 1 (Sec Model input): For UE-sided models (Case 1/2a), when both training and inference are performed on the UE side, the model input format is determined by the UE vendors
- Proposal 5 (Sec Model input): Support Option B for the starting time of the measurement window - the starting time of the list of Nt consecutive samples is the timing of the earliest sample that the first path power is detectable
- Proposal 8 (Sec Model output): In the AI/ML-assisted positioning method, the timing information output from the model should be used independently, without incorporating the LOS/NLOS indication
- Proposal 9 (Sec Model output): In the AI/ML-assisted positioning method, the LOS/NLOS indicator output from the model should be incorporated with the legacy timing measurement
- Proposal 11 (Sec Quality indicator): For PDP channel measurement, L1-SINR can serve as a quality indicator
- Proposal 14 (Sec Time stamp): For the time stamp of measurement and label, NR-TimeStamp-r16 from TS37.355 with the addition of UTCTime and TimeStamp in TS38.455 can be reused
- Proposal 15 (Sec Association of Part A and Part B): For training data collection, an additional tag is required for both Part A and Part B to ensure efficient dataset classification
- Proposal 16 (Sec Consistency): For the AI/ML positioning model, ensure consistency between training and inference by adopting the associated-ID
- Proposal 17 (Sec Label-based performance monitoring): For label-based model performance monitoring of AI/ML positioning Case 1, Option A is preferable to Option B
- Proposal 18 (Sec Label-based performance monitoring): Option A-3 is the most feasible and reliable solution for monitoring in Case 1
- Proposal 21 (Sec Label-free model performance monitoring): For label-free performance monitoring in direct AI/ML models, consider using legacy measurements and assisted positioning model outputs, in addition to leveraging model input/output statistics
- Proposal 6 (Sec Model input): For the definition of sample-based measurement, if bandwidth aggregation is applied, the set of candidate values for k include [-6, 5]
- Proposal 3 (Sec Model input): For gNB-sided model (Case 3a), when both training and inference are performed on the gNB-side, the model input format is determined by the gNB manufacturers
- Proposal 19 (Sec Label-based performance monitoring): For label-based model performance monitoring of AI/ML positioning Case 1, further study how to assess the reliability of approximate ground truth labels
- Proposal 20 (Sec Label-based performance monitoring): For Case 3a, in the absence of a PRU, explore methods to assess the reliability of ground truth labels used in label-based performance monitoring