Anh-Quan Pham

We investigate whether large language models and massive compute can ease the challenge of training RL-based manipulation agents for zero-shot sim-to-real transfer. By integrating LLMs into the control loop and leveraging hundreds of Nvidia’s latest-generation data-center GPUs, we explore a new frontier in scaling robotic generalization.

We introduce a conditional diffusion model that synthesizes full RL transitions for unseen tasks using compositional task indicators. Trained on a subset of expert data, it enables zero-shot policy learning with synthetic replay and solves novel task compositions through online RL by recombining subskills from seen tasks.

We design an intelligent traffic steering system for Open RAN by jointly optimizing flow-split, congestion, and scheduling across layers. Our layered rAPP–xAPP solution enables adaptive, UE-aware control via network utility maximization under delay and stability constraints.