A novel voice in head actor critic reinforcement learning with human feedback framework for enhanced robot navigation

Area of Science:

• Robotics
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Robotic navigation and interaction in complex environments remain challenging.
• Integrating natural language understanding with robotic control is an active research area.
• Current systems often lack adaptability and sophisticated reasoning capabilities.

Purpose of the Study:

• To introduce a novel Voice in Head (ViH) framework for enhanced robotic navigation and interaction.
• To leverage Large Language Models (LLMs) and semantic search for intuitive human-robot communication.
• To develop a safe, adaptable, and scalable cognitive robotics system.

Main Methods:

• Integration of GPT and Gemini powered LLMs as Actor and Critic components within a reinforcement learning (RL) loop.
• Utilization of Azure AI Search for a semantic search mechanism enabling natural language queries.
• Implementation of Reinforcement Learning with Human Feedback (RLHF) for safety and addressing LLM limitations.

Main Results:

• The ViH framework achieved success rates of up to 94.54%, outperforming existing benchmarks.
• Demonstrated effective robotic navigation and interaction in complex environments.
• The system proved modular and scalable, adaptable to diverse application domains.

Conclusions:

• The ViH framework represents a significant advancement in cognitive robotics.
• The hybrid approach combining LLMs, RL, and semantic search enhances autonomous system capabilities.
• This research contributes to the development of intelligent autonomous systems and progresses towards Artificial General Intelligence.