Framework for feedback elicitation in virtual-reality-based reinforcement learning from human feedback

Kurzfassung

Modern manufacturing is increasingly adopting collaborative robots, motivated by improvements in productivity and flexibility. For the safe, secure, and effective implementation of AI control algorithms for robotic workers, a careful training process is required. Reinforcement Learning from Human Feedback (RLHF) is a valuable tool for realising this training process. In this paper, we present a framework for RLHF based on Virtual Reality (VR) that supports designing virtual training environments for robotic workers. We also built a VR environment that demonstrates a practical implementation of the framework, focusing on the elicitation of human feedback. We detail key implementation aspects, including the general system design and the design of the feedback interface. Finally, we evaluate the VR environment through a user study involving 48 participants.