Sightless Showdown: a Comparative Study of AI and Human Performance Without Vision in a Fighting Game

This paper presents a comparative evaluation of a blind reinforcement learning agent and visually impaired human players in a fighting game using only audio inputs. The study employs DareFightingICE, a fighting game platform with enhanced sound design tailored for visually impaired users, to assess whether a blind AI trained with audio inputs can perform comparably to domain-expert human players under sightless conditions. The blind Artificial Intelligence (AI), trained over 900 rounds using three audio encoders (1D convolutional neural networks, fast Fourier transform, and Mel-spectrogram) and data from 25 domain-expert human players (21 male, 3 female, 1 non-binary; age range = 18–35 years), was tested against a weakened Monte Carlo Tree Search opponent. Results indicate no statistically significant difference in performance between the blind AI and human players, suggesting that audio-driven AI can serve as a viable proxy for human testing during early-stage sound design validation in accessible game development.