OpenNetLab: An Open Platform for AI-Based Bandwidth Estimation in Real-Time Communications
Abstract
Real-time communications (RTC), such as video conferencing and cloud gaming, rely on bandwidth estimation (BWE) to balance throughput, delay, and loss under diverse network conditions and media workloads. AI-based BWE is a promising direction beyond hand-tuned rules, but progress remains slow and fragmented because each project must rebuild the system support around the model itself: a real RTC stack for running models; environments for training, validation, and real-world evaluation; and artifacts for comparing and improving models. This paper presents OpenNetLab, an open platform for AI-based BWE in RTC. At its core, OpenNetLab provides simulator, emulator, and real-node runtimes for training, validation, and real-world evaluation. The three runtimes share AlphaRTC, which combines the WebRTC stack with a common AI-BWE interface, while runtime-specific Media and Network Engines meet the workflow’s efficiency and fidelity requirements through fast simulation, real-stack emulation, and live-Internet execution. A shared Benchmark defines reproducible cases and metrics, while Open Artifacts support independent deployment, community sharing and contributions, and continual benchmark updates. We evaluate OpenNetLab through matched-replay execution fidelity, training efficiency, and benchmark comparability. A 600-case matched-replay study shows small gaps in network and media outcomes between the OpenNetLab emulator and a WebRTC reference. A PPO training run compresses about 194 call-hours into 23 minutes of simulator time, a roughly 497 × speedup. OpenNetLab has powered the MMSys 2021 BWE Challenge and has been used in subsequent published RTC research.