• ai
• architecture
• artificial intelligence
• cognitive
• explainable ai
• large language models
• llm
• mqtt
• robotics
• service
• social

My thesis focuses on developing and testing a cognitive architecture for a humanoid robot, embodied by Abel, from the University of Pisa. Abel can express emotions, speak, see, and hear using advanced sensors and servomotors.

Previous architectures relied on rule-based systems, Markov Decision Processes, or reinforcement learning to approximate human mental models. The emergence of Large Language Models (LLMs) has provided a new method for this task.

My thesis uses an LLM within a chain framework (Microchain) to manage function calls among predefined functions: Look, Hear, Talk, Recognize, Recall, Memorize, Emotion, Reasoning, and Stop. This setup is inspired by neuroscience studies showing brain regions specialized for specific functions, including high-level cognitive tasks.

The architecture is service-oriented, ensuring decoupling and fault tolerance. Communication between Abel and the services occurs via MQTT, with both synchronous and asynchronous calls.

Validation involves process mining tools, specifically transition maps. Experiments with 20 participants tested four LLMs in a predefined, ambiguous scenario to assess function calls and LLM behaviours.

Results showed similar transition maps, with Llama70b performing double the 'Look' calls, enhancing situational awareness and preference among participants (18 over 20). Future improvements could include using faster protocols like RabbitMQ, comparing other LLMs, and incorporating psychological tests.