Imagine robots seamlessly navigating our homes, assisting in complex tasks, and even performing intricate manufacturing processes—all thanks to the power of Artificial Intelligence. This future hinges on the efficient integration of powerful AI models with the physical constraints of robots in the real world. Large Language Models (LLMs), known for their exceptional reasoning and planning abilities, are playing a pivotal role in this evolution. However, there's a catch. Current AI-powered robots operate on a frame-by-frame basis, where they process visual input and generate actions sequentially. This process creates a bottleneck, causing significant delays as the AI, robot control systems, and data communication struggle to keep pace. This paper introduces 'Corki,' an innovative framework that tackles these challenges head-on. Corki's brilliance lies in decoupling these three crucial elements: LLM inference, robotic control, and data communication. Instead of predicting actions for single frames, Corki predicts the robot's trajectory for the near future. This reduces how often the AI needs to process information, saving valuable time and energy. This algorithmic leap is coupled with a specialized hardware accelerator that swiftly translates the AI's predicted trajectory into real-time control signals for the robot's motors. Furthermore, Corki cleverly overlaps data transmission with the robot's execution process, further minimizing delays. Results show Corki achieves an impressive speed improvement, making robots up to 3.6 times faster! It also boosts their success rate in complex tasks by up to 17.3%. This breakthrough demonstrates the transformative potential of a holistic approach, co-designing software and hardware for smarter, more efficient AI robots.