Adapting Virtual Embodiment Through Reinforcement Learning.

Area of Science:

• Virtual Reality (VR)
• Human-Computer Interaction
• Perception

Background:

• Virtual bodies in VR enable diverse applications like perspective-taking and enhanced training.
• Avatar movement distortion can improve user experience but may break embodiment if excessive.
• Individual sensitivity to movement distortion varies significantly.

Purpose of the Study:

• To propose and validate a Reinforcement Learning (RL) method for accurately determining an individual's distortion detection threshold.
• To compare the robustness and noise-handling capabilities of the RL method against the adaptive staircase method.

Main Methods:

• Utilized Reinforcement Learning (RL) to dynamically adjust and identify the maximum unnoticed avatar movement distortion.
• Conducted a controlled experiment with human subjects to collect data on distortion perception.
• Implemented a majority voting system within the RL framework to manage noisy input data.

Main Results:

• The RL method determined a more robust detection threshold than the adaptive staircase method.
• The RL approach was more effective at preventing subjects from detecting distortions below their identified threshold.
• The RL method demonstrated superior ability to handle noisy input data compared to the adaptive staircase.

Conclusions:

• Reinforcement Learning offers an efficient and robust approach to calibrate individual embodiment in VR by precisely identifying distortion thresholds.
• The RL method's noise-handling capacity is crucial for future applications using physiological signals for real-time calibration.
• Personalized embodiment calibration can significantly enhance the effectiveness of VR interactions for training, rehabilitation, and other applications.