Kenza: Embodied Multimodal Sentient Companion

Abstract

Although Large Language Models (LLMs) and multimodal AI are quickly expanding the capabilities of embodied agents, real world robotic systems continue to face major challenges: the high cost of hardware, inflexible interaction models, and excessive dependence on cloud computing. We will address these challenges by proposing Kenza, an embodied AI companion that will promote a smooth and natural interaction between humans and robots. Kenza is an open-source platform that combines computer vision, speech processing, and LLM driven reasoning into a single platform, powered by an accessible Raspberry Pi 5. We use a strategic hybrid intelligence pipeline. We delegate multifaceted conversational reasoning to cloud based LLMs, and time sensitive tasks like vision, gesture recognition and navigation, run on the local edge device only. This bifurcated architecture ensures that the robot is very responsive even when there is a fluctuation in network connectivity. We show that advanced interactive features such as voice based dialogue, gesture control, and automatic user tracking can be fully implemented on low cost hardware. Finally, Kenzo provides low latency performance, high environmental perception, and consistent indoor navigation, which forms a highly practical basis of future studies in social robotics and embodied AI.