R1-2409669
discussion
Specification support for positioning accuracy enhancement
From vivo
Summary
This document from vivo analyzes specification impacts for AI/ML-based positioning in NR, focusing on data collection, model inference, and consistency between training and inference. It presents simulation results demonstrating the superiority of sample-based measurements over path-based measurements and proposes specific configurations for reporting, phase information usage, and model monitoring to ensure positioning accuracy.
Position
vivo argues that sample-based measurements significantly outperform path-based measurements for AI/ML positioning, particularly in scenarios with limited TRPs or bandwidth, and proposes specifying sample-based reporting with defined parameters (Nt, Nt’, k). They prefer reusing existing IEs for quality indicators and assistance data to minimize specification overhead, while suggesting implicit indication via associated IDs for privacy-sensitive information. vivo supports the inclusion of phase information (RSCP/RSCPD) and distance ranging as model inputs/outputs to enhance accuracy. They emphasize the critical need for consistency between training and inference, proposing associated IDs to manage NW-side conditions like beam patterns and bandwidth alignment. Finally, they propose defining Case 1 as a new positioning method to facilitate proper assistance data handling and model monitoring procedures.
Key proposals
- Proposal 4.1.5-1A (Sec 2.1): Proposes reusing existing IEs (confidence and uncertainty) for label quality indicators rather than defining new abstract IEs, arguing existing mechanisms are sufficient.
- Proposal (Sec 2.1): Proposes reusing NR-TimingQuality and Timing Measurement Quality IEs for channel measurement quality, associating one quality indicator with the entire channel measurement rather than individual paths.
- Proposal (Sec 2.1): Suggests identifying necessary assistance information for model training that should be mandatorily provided by LMF, with privacy-sensitive data implicitly indicated via an associated ID.
- Proposal (Sec 3.1): Proposes specifying sample-based channel measurement reporting for Case 2b and 3b, arguing it provides better positioning performance than path-based measurement, especially in limited TRP/bandwidth scenarios.
- Proposal (Sec 3.1): Proposes supporting Option A for determining the starting time of Nt consecutive samples, where the start time is the timing of the first detected path.
- Proposal (Sec 3.1): Proposes candidate values for sample-based measurement parameters: Nt in {32, 64, 128}, Nt’ in {8, 16, 32}, and k in {3, 4, 5}.
- Proposal (Sec 3.1): Proposes supporting RSCP and RSCPD for AI/ML positioning to utilize phase information, which can offer at least 20% accuracy gain over PDP with negligible overhead.
- Proposal (Sec 3.1): Proposes supporting distance ranging between UE and TRP for Case 2a and 3a model outputs, which provides over 45% accuracy gain compared to DL-RSTD.
- Proposal (Sec 3.1): Proposes attaching an indicator to timing measurements to specify if they are generated by AI/ML models, or alternatively, not reporting LOS/NLOS indicators for AI/ML-generated timing.
- Proposal (Sec 4.1): Proposes supporting both label-based and label-free model monitoring, with label-based monitoring performed at the target UE side using ground truth labels from LMF or PRU information.
- Proposal (Sec 5.1): Proposes adopting associated IDs to ensure consistency of NW-side conditions (e.g., beam patterns) between training and inference, connecting IDs to cell/TRP lists.
- Proposal (Sec 5.1): Proposes that UE report supported PRS bandwidth to ensure consistency of measurement bandwidth between training and inference.
- Proposal (Sec 5.2): Proposes specifying AI/ML based positioning Case 1 as a new positioning method to align with RAN2 procedures and handle unique assistance data requirements.