R1-2409925
discussion
Specification support for AI/ML-based beam management
From CATT
Summary
CATT presents a comprehensive technical contribution on AI/ML-based beam management for 5G NR Rel-19, covering both UE-sided and network-sided models with 39 detailed proposals addressing configuration, inference, reporting, and performance monitoring aspects.
Position
CATT advocates FOR flexible AI/ML beam management solutions that reuse existing CSI framework mechanisms while introducing minimal specification impact. They push FOR supporting both UE-sided and network-sided models with comprehensive performance monitoring, AGAINST overly complex new signaling mechanisms, and favor practical approaches like reusing legacy TCI delay requirements and CPU mechanisms rather than creating entirely new frameworks.
Key proposals
- Proposal 1 (Sec 2.1): Conclude that for UE sided model in beam management, no need to define the similar properties of a DL Tx beam or beam set/list with the associated ID
- Proposal 5 (Sec 2.2): For NW-sided model, it is beneficial to align the Rx information of the measurements between network and UE
- Proposal 7 (Sec 3.1.1): For inference of UE-sided model, at least for BM-Case1, support Alt 1 (one CSI-ResourceConfigId for Set B with Set A from associated ID) and Alt 3 (two separate CSI-ResourceConfigIds for Set A and Set B)
- Proposal 8 (Sec 3.1.2): For UE-sided AI/ML model inference, for BM-Case2, AP CSI-RS can be supported for CMR configuration at least for Set B
- Proposal 13 (Sec 3.2): For UE-sided model, at least for BM-Case1, the reported beam information can be RS indicator(s) or pre-defined beam index of predicted Top K beam(s)
- Proposal 17 (Sec 3.3): From RAN1 perspective, the existing TCI state switching delay requirements can be reused for the predicted RS resource of UE sided-model
- Proposal 20 (Sec 3.4.2): For BM-Case1 and BM-Case2 with UE-sided AI/ML model, support Top K beam prediction accuracy metrics, L1-RSRP difference information, and RSRP difference between predicted and measured values
- Proposal 29 (Sec 3.5): Further study how to measure predicted Top-K beams in CSI framework via extending trigger states or introducing dynamic QCL relations indication
- Proposal 30 (Sec 3.6): For UE side AI/ML-based CSI processing, a new type of processing unit can be introduced if different modules are used for RS measurement and model inference
- Proposal 32 (Sec 4.1): For NW-sided model, support L1-RSRPs and beam information of up to M beams within X dB gap to largest measured L1-RSRP value
- Proposal 36 (Sec 4.1): At least for NW-sided model, support introducing larger quantization step size for differential L1-RSRP reporting
- Proposal 37 (Sec 4.2): For data collection training for NW-sided model, support L1-RSRPs of Set A/B or L1-RSRPs of Set B plus Top K beam information of Set A