Research
Programmable O-RAN and MEC
Open radio control and edge computing turn AI decisions into coordinated radio, slicing, and offloading actions.
Original figure from the FACTORY-AI proposal.
Hardware-aware AI execution
3D heterogeneous integration and compute-in-memory connect efficient execution to the project’s adaptive control loop.
Original figure from the FACTORY-AI proposal.
A factory that senses, decides, and adapts
FACTORY-AI unifies perception, communication, and computing in a closed control loop. Sensor information becomes actionable semantic tokens; AI-native network control allocates radio and edge resources; heterogeneous hardware executes the workload and returns feedback that improves the next decision.
Thrust 1 • Perception–Learning–Reflex
Make sensing actionable. ISAC-GPT turns integrated sensing and communication observations into task-relevant control tokens. Intent-aware federated learning combines reinforcement learning with over-the-air aggregation, while distributed near-real-time RIC reflexes respond to local conditions.
The aim is to preserve what a factory decision needs while reducing the cost of transmitting and processing unnecessary data.
Thrust 2 • O-RAN edge intelligence
Orchestrate the network and the edge together. LLM-augmented reinforcement learning coordinates radio slices, offloading, and MEC resources through an open, programmable O-RAN stack. Predictive policies anticipate changing traffic and balance end-to-end service requirements.
Private 5G, OpenAirInterface, RIC applications, and joint communication–compute scheduling connect the intelligence to executable network actions.
Thrust 3 • Heterogeneous computing infrastructure
Execute AI efficiently under real hardware constraints. The project explores 3D heterogeneous integration and compute-in-memory, token-aware mixture-of-experts mapping, mixed precision, and adaptive runtime scheduling.
Execution feedback on latency, energy, and resource pressure returns to both perception and orchestration, keeping the entire system responsive to changing workloads.