R1-2409878
discussion
Discussion on AI/ML-based positioning accuracy enhancement
From Xiaomi
Summary
Xiaomi presents a comprehensive discussion on AI/ML-based positioning accuracy enhancement for NR, covering input types, data collection, functionality identification, inference, and performance monitoring across different positioning cases. The document contains 28 proposals and 2 observations addressing key aspects of AI/ML positioning standardization.
Position
Xiaomi strongly advocates FOR sample-based input over path-based input for Cases 2b and 3b due to better positioning accuracy and reduced ambiguity issues, while pushing AGAINST CIR support due to high overhead with marginal performance gains. They favor UE-side performance monitoring and LCM decision making to reduce latency, and support explicit indication of assistance data rather than implicit approaches. Xiaomi emphasizes implementation flexibility for UE-side cases while requiring standardization for network-side cases.
Key proposals
- Proposal 1 (Sec 1): For Case 2b and Case 3b, support sample-based input. For Case 1, Case 2a and Case 3a, it is up to implementation to determine the model input
- Proposal 5 (Sec 1): The support of CIR should be deprioritized
- Proposal 6 (Sec 2.1): Confirm working assumption for Case 1 that measurement and related data are generated by PRU and/or Non-PRU UE, with labels generated by PRU, Non-PRU UE with estimated location, or LMF
- Proposal 8 (Sec 2.2): Consider two options as starting point for reference time of measurement report in Case 2b - reuse UL-RTOA definition or utilize receiving sub frame boundary
- Proposal 11 (Sec 2.3): For Case 3a, leave RAN3 to assess specification impact on collected data format definition and information exchange between TRP/gNB and different gNBs
- Proposal 12 (Sec 3): Support functionality identification for Case 1 and Case 2a
- Proposal 14 (Sec 3): Categorize AI/ML feature from 3 aspects - output type of AI models, deployment location, and positioning RS type (SRS-based or PRS-based)
- Proposal 16 (Sec 4.1): In Case 1 and Case 2a, there is no specification impact on delivery of input data for inference or inference output data
- Proposal 18 (Sec 4.3): For Case 3b and Case 2b, input data format for inference needs to be specified and common format can be used for both data collection and inference
- Proposal 19 (Sec 5.1): Utilize explicit indication manner to indicate existing assistance data E-based DL-TDOA and/or UE-based DL-AoD
- Proposal 21 (Sec 6.1.1): For Case 1 performance monitoring, performance metric calculation and LCM decision making can be performed on UE side
- Proposal 24 (Sec 6.1.3): For Case 3a input-based monitoring, gNB calculates metric and makes LCM decision for multi-TRP construction, while LMF makes decision for single TRP construction
- Proposal 26 (Sec 6.2): Confirm necessity of assessment/monitoring of inactive models/functionalities with various approaches including monitoring-only mode and dataset delivery
- Proposal 27 (Sec 7): The fallback mode including non-AI positioning approach and related configuration are preconfigured to UE
- Proposal 28 (Sec 7): UE group-based LCM operation (model activation/deactivation/switch) can be considered