Anh-Quan Pham

We leverage compositional structure in robotic manipulation to iteratively generate expert data for unseen task combinations via conditional diffusion. A semantic compositional diffusion transformer factorizes task components and self-improves through offline RL validation, enabling strong zero-shot data generation to solve new tasks without collecting new real-world data.

We introduce a system for automatically constructing reinforcement-learning–ready environments from high-level human specifications. By combining a scalable manipulation environment backbone, LLM-generated rewards with evolutionary refinement, and a VLM-based articulated object refinement pipeline, EurekaWorld enables large-scale policy training with realistic physics and substantially reduced manual environment and reward engineering.

SBAMP combines RRT* for global path planning with an online SEDS-based controller for real-time local trajectory adaptation. The method avoids any offline training data by fitting stable dynamical systems on-the-fly and ensures global Lyapunov stability during disturbances and replanning.

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.