This project aims to develop a real‑time embodied AI system on mobile manipulation platforms (e.g., dual‑arm robots or mobile bases equipped with manipulators) to support tasks such as object search, pick‑and‑place, and human–robot interaction in dynamic indoor environments. The key focus is on optimizing the end‑to‑end perception–reasoning–action pipeline of Vision–Language–Action (VLA) and World‑Action Models (WAMs) for efficient on‑device deployment. We will investigate techniques including pipelined sensing, cross‑modal early exit, and adaptive model configuration to achieve low‑latency and resource‑efficient execution on edge platforms (e.g., NVIDIA Jetson). The developed system will be validated on real robotic hardware in dynamic settings, enabling continuous robot navigation, manipulation, and interactive behaviors under real‑world constraints.