R1-2500466
discussion
On specification for AI/ML-based positioning accuracy enhancements
From OPPO
Summary
OPPO submits 33 proposals and 3 observations regarding specification impacts for AI/ML-based positioning accuracy enhancements in Rel-19, covering measurement enhancements, training/inference consistency, data collection, model inference, and performance monitoring. The document argues against supporting phase information reporting and dedicated specification enhancements for monitoring without ground-truth labels, while proposing the use of an 'associated ID' to ensure consistency between model training and inference.
Position
OPPO opposes supporting reporting based on phase information for Rel-19 AI-based positioning, arguing that timing and power information are sufficient and phase adds unjustified overhead. They propose using an 'associated ID' signaled from the network to ensure consistency between AI model training and inference for UE-side models (Case 1 and 2a), rather than relying solely on validity areas which fail to account for temporal network changes. OPPO argues that Rel-19 should NOT specify dedicated enhancements for performance monitoring without ground-truth labels, leaving such mechanisms to implementation. For data collection, they propose reusing legacy LPP/NRPPa signaling without specifying data formats, asserting that content is up to implementation. They further propose that the network should not specify mechanisms to deliver training data between different entities, and that UE/gNB should indicate whether results are AI-based or legacy-based in reporting.
Key proposals
- Proposal 1 (Measurement enhancement): The set of candidate values for timing reporting granularity factor k include at least {0, 1, 2, 3, 4, 5}.
- Proposal 2 (Measurement enhancement): The set of candidate values for Nt’ (number of samples) include at least {9, 16, 24}.
- Proposal 4 (Measurement enhancement): For R19 AI-based positioning, NOT support the reporting based on phase information in addition to timing and power information.
- Proposal 6 (Consistency between training and inference): Signal an indication (e.g., associated ID) from the network to ensure consistency of AI model training and inference for Case 1 and Case 2a, without disclosing proprietary network information.
- Proposal 10 (Data collection for training): Reuse legacy LPP signaling to configure UE with positioning RS for data collection; no additional signaling is needed to start collection, and data format/content is up to UE implementation.
- Proposal 14 (Data collection for training): Rel-19 is NOT to specify any mechanism to deliver collected data from the entity obtaining training data to the training entity when they are different entities.
- Proposal 17 (AI model inference and reporting): For Case 1, introduce new information to indicate whether reported results are generated by legacy method or AI-based method; AI model input format is up to UE implementation.
- Proposal 21 (AI model inference and reporting): For Case 3a, support combinations where timing and LOS/NLOS indicators are generated by either AI/ML or legacy methods independently.
- Proposal 24 (Functionality/model performance monitoring): Rel-19 is NOT to specify dedicated specification enhancement for functionality/model performance monitoring mechanisms without ground-truth labels.
- Proposal 29 (Functionality/model performance monitoring): Support enhancement on NRPPa signaling to enable delivery of ground-truth labels from LMF to gNB for Case 3a monitoring.
- Proposal 31 (Functionality/model management): UE can autonomously deactivate AI operations and fall back to legacy operations, with reporting indicating the source of the result.
- Proposal 33 (Functionality/model identification): UE can report applicable functionalities by sending a ProvideCapabilities message to the LMF.