AI-TPC on O-RAN: A Safe and Deployable Framework with OTA Validation on OCUDU

Abstract

This paper presents a safe and deployable AI-based PUSCH transmit power control (AI-TPC) framework for O-RAN, implemented inside the per-slot scheduling loop of an open-source O-DU (OCUDU release_26_04). The framework supports two inference modes: a supervised MLP-TPC baseline and an offline reinforcement-learning CQL-TPC policy, both protected by a four-condition runtime fallback to the legacy 3GPP accumulated TPC (ACC) path. The models are deployed with ONNX Runtime on CPU only, achieving P99 inference latencies of ${28.5} \mu \mathrm{s}$ and ${31.0} \mu \mathrm{s}$, respectively, while preserving the existing signaling flow. We validate the framework over the air on a real O-RU and commercial UE under four target SINR settings from 26 to 29 dB. Results show that MLP-TPC validates the deployability of the ONNX inference pipeline inside the O-DU scheduler, while CQL-TPC improves power-headroom efficiency by 1.6-3.8 dB without significant throughput or BLER loss. These results demonstrate that AI-based TPC can be integrated into an O-RAN scheduler in a practical and safety-aware manner.