Research
-
mmsRAN: Multimodal Cellular Uplink for Physical AI
Robots stream LiDAR, camera, and inertial data over the cellular uplink to an edge application that closes their control loop, and every stream for a timestep must arrive in time for the next action. Uneven arrival across sensors and users adds significant latency. mmsRAN explores coflow-inspired uplink scheduling that delivers each timestep's sensor streams together, within and across robots.
-
Scout: User Plane Telemetry for NextG Cellular Networks
Scout is packet-centric telemetry for the 5G user plane. It traces packets through the RAN stack with low overhead, capturing per-packet latency, queue dynamics, and the impact of scheduling decisions. Integrated with EdgeRIC, it feeds ML-based μApps and provides ground truth for evaluating and debugging control policies.
-
Windex: A 5G scheduler for scalable service guarantees
Windex is a scalable 5G scheduler that gives each user guarantees on throughput and service regularity. It uses neural networks to compute a Whittle index per user and allocates resources in order of these indices, turning a combinatorial problem into a linear one. Running on a real-time RIC, it makes decisions in under 20 μs per user and significantly improves service guarantees over existing schedulers.
-
SPARC: Multi-site Spectrum Management
SPARC manages limited licensed spectrum across multiple cell sites, using base stations as spectrum sensors. An xApp detects and localizes interference in near-real time, and a μApp allocates resources in real time. SPARC improves SNR by up to 7 dB, spectral efficiency by up to 15%, and system throughput by up to 20%.
-
Tiny-Twin: A CPU-Native Full-stack Digital Twin for NextG Cellular Networks
Tiny-Twin is an open-source, full-stack 5G digital twin that runs on commodity CPUs, with no specialized hardware. It combines a complete 5G protocol stack with time-varying multi-tap channel convolution, per-UE channel isolation, channel trace replay, and built-in real-time RIC integration. It scales to multiple UEs while preserving protocol timing, offering a middle ground between low-fidelity simulators and costly hardware emulators.
-
EdgeRIC: Delivering Realtime RAN Intelligence
EdgeRIC is a real-time RAN intelligent controller co-located with the Distributed Unit. It is decoupled from the RAN stack but operates at RAN timescales (sub-millisecond), running AI-in-the-loop μApps that use RAN and application-level information to make decisions. EdgeRIC outperforms cloud-based near-real-time RICs, and its AI policies beat model-based strategies by 5% to 25% in throughput and application-level metrics.
