R1-2500405
discussion
Specification support for positioning accuracy enhancement
Summary
Tejas Networks Limited presents 24 proposals and 15 observations regarding AI/ML for NR positioning accuracy, focusing on sample-based and path-based measurement inputs, model output definitions for Case-3a, training data collection frameworks, and model performance monitoring. The document argues for reusing existing positioning frameworks while introducing specific enhancements for AI/ML contexts, such as redefining LoS/NLoS indications and standardizing sample window parameters.
Position
Tejas Networks proposes specific parameter sets for sample-based measurements, including Nt values of {32, 64, 128} and Nt' as a fraction of Nt, to balance positioning accuracy and reporting overhead. They require redefining the LoS/NLoS indication for Case-3a to reflect the likelihood that the reported timing corresponds to the direct path, rather than the legacy probability of link existence, arguing that legacy indicators do not improve accuracy in NLoS terrains. They support reusing existing Release-17/18 positioning frameworks for training data collection and assistance data, asserting that current IEs are sufficient with minor enhancements like validity area context. For Model Performance Monitoring, they propose consolidating metric calculation and decision-making at a single entity (LMF or UE/TRP) to minimize specification impact, using range error or positioning error as the primary metrics. They also support implicit reporting of TRP geographical coordinates to protect sensitive location data during the inference phase.
Key proposals
- Proposal 1 (Model Input - Sample-based): Supports Nt values of {32, 64, 128} for the size of the sample window based on deployment scenarios.
- Proposal 2 (Model Input - Sample-based): Proposes that the number of reported samples (Nt') should be a fraction of the window size (Nt), selected by the TRP or LMF.
- Proposal 9 (Model Input - Path-based): Utilizes the existing Release-17 framework for path-based measurements, potentially increasing the number of reported paths beyond 8.
- Proposal 10 (Model Output - Case-3a): Proposes that timing information reported by the gNB should correspond to the ToA of either the actual or inferred non-existent LoS path.
- Proposal 11 (Model Output - Case-3a): Proposes a new LoS/NLoS indication reporting the likelihood that the reported timing information corresponds to the direct/LoS path, rather than just link existence.
- Proposal 12 (Model Output - Case-3a): Argues against inferring legacy LoS/NLoS indications using AI/ML models for reporting with legacy timing information.
- Proposal 14 (Training Data Collection): Proposes associating multiple Part-A measurements to a single Part-B label for direct AI/ML positioning when reported by a single entity.
- Proposal 17 (Training Data Collection - Case-1): States that the existing positioning framework is sufficient for collecting training data for UE-based positioning.
- Proposal 19 (Model Inference - Consistency): Supports Alternative-1 for implicit reporting of TRP geographical coordinates (Info #7) to protect sensitive information.
- Proposal 22 (Model Performance Monitoring): Proposes using range error or positioning error as the model performance metric for decision-making.
- Proposal 23 (Model Performance Monitoring): Proposes that both MPM metric calculation and decision-making should occur at a single entity (LMF or UE/TRP) to reduce specification impact.
- Proposal 24 (Model Performance Monitoring): Proposes periodic or aperiodic monitoring of AI/ML model performance due to potential degradation from mobility or environmental changes.