Human-AI-Robot Collaboration : Designing conceptual HARC framework with reinforcement learning for unstructured environments

Abstract

This thesis introduces the conceptual Human-AI-Robot collaboration (HARC) framework. The core purpose behind this research is to validate technologies for the conceptual HARC framework to work as a fleet of robots in real-world unstructured environments under human supervision. The current issue in the real-world environments is the traditional reinforcement learning that relies on static internet-based datasets, trial-and-error and episodic feedback. In addition, these RL agents struggle in real-world unstructured environments as they fail to generalize and adapt due to real worlds changing dynamic outcomes.

This thesis investigates technologies with potential to overcome these limitations. Technologies include human-guided reinforcement learning (HG-RL), edge AI, neuromorphic computing and federated learning (FL). HARC intends to collaborate these technologies into one scalable modular architecture. Central to this framework is the human role as an active policy shaper who supervises multiple agents in real time to maximize productivity.

In addition, the supervisor operates multi-monitor setups, providing feedback, allowing efficient behavior correction, and policy optimization. A practical application scenario illustrates a human supervising multiple neuromorphic tractors, using HARC to offer context-specific feedback and generate real-time policy updates without unreliable cloud infrastructure. HARC transforms reinforcement learning into a continuous process of agent evolution through human insight and real-world interaction, positioning humans as active operators.