R1-2500636
discussion
Specification impacts for AI/ML positioning
From Lenovo
Summary
Lenovo submits 28 proposals and 1 observation to advance the specification of AI/ML-based positioning in NR, focusing on sample-based measurement definitions, model input/output types, training data construction, and model management. The document aims to resolve open issues regarding timing granularity, channel impulse response (CIR) reporting, and the reuse of legacy signaling frameworks for AI/ML model training and monitoring.
Position
Lenovo proposes extending sample-based measurement definitions to UE-based Cases 1 and 2b, requiring the definition of reference time and parameters {Nt, Nt’, k}. They support the inclusion of DL and UL Channel Impulse Response (CIR) measurements, as well as legacy measurements like RSTD and RSRP, as model inputs for fingerprinting. Lenovo requires the reuse of legacy LOS/NLOS indicator frameworks for AI/ML model outputs and supports Multi-RTT timing measurements for LMF-side models. They propose introducing a new UE-based positioning method for Direct AI/ML Case 1 and support Alternative 2 for signaling TRP geographical coordinates (Info #7). For training data, Lenovo supports indications for labelled versus unlabelled data samples and confirms working assumptions that labels can be generated by PRUs, Non-PRU UEs, or the LMF. Finally, they propose reporting model monitoring outcomes based on positioning accuracy and introducing mechanisms for UEs to retrieve ground truth labels from the LMF.
Key proposals
- Proposal 1 (Sec 2.1.1): Extend sample-based measurement definitions to Cases 1 and 2b, with FFS on reference time, {Nt, Nt’, k} values, and UE capabilities.
- Proposal 4 (Sec 2.1.2): Conclude on additional path-based enhancements, such as extending the maximum number of reported additional paths.
- Proposal 5 (Sec 2.1.3): Support sample selection rules based on minimum and maximum path power to reduce measurement overhead.
- Proposal 6 (Sec 2.2): Support DL Channel Impulse Response (CIR) measurements as model input for DL-based Direct AI/ML positioning.
- Proposal 7 (Sec 2.2): Support UL CIR measurements and UL-based angle-delay domain profiles as model input for UL-based Direct AI/ML positioning.
- Proposal 8 (Sec 2.2): Re-use legacy positioning measurements (e.g., DL-RSTD, UL-RTOA, RSRP) as model input types to derive fingerprints.
- Proposal 9 (Sec 2.3): Re-use the legacy LOS/NLOS indicator reporting framework for AI/ML model outputs.
- Proposal 11 (Sec 2.4): Support paired gNB Rx-Tx and UE Rx-Tx time difference measurements (Multi-RTT) for LMF-side models in Case 3b.
- Proposal 13 (Sec 3.1.2): Define ground truth label content for each case, including location information, uncertainty, and timestamps.
- Proposal 15 (Sec 3.1.3): Support indications for providing only Part A (unlabelled) or Part A and Part B (labelled) of training data samples via LPP/NRPPa.
- Proposal 16 (Sec 3.2.2): Introduce a new UE-based positioning method specifically for Direct AI/ML Positioning Case 1.
- Proposal 19 (Sec 3.2.3): Specify data request/collection scenarios for LMF-side, UE-side, and gNB-side model training, considering non-3GPP technologies for UE/gNB training.
- Proposal 23 (Sec 4.1): Report model monitoring outcomes per AI/ML model based on horizontal/vertical positioning accuracy.
- Proposal 25 (Sec 4.1): Introduce a signaling mechanism for UEs to request and retrieve ground truth label information from the LMF for Case 1 monitoring.
- Proposal 28 (Sec 4.3): Further study mechanisms for efficient positioning AI/ML model transfer between UE, gNB, and LMF.