Assimilating human feedback from autonomous vehicle interaction in reinforcement learning models.

Area of Science:

• Human-computer interaction
• Robotics
• Behavioral science

Background:

• Interactions between automated vehicles (AVs) and pedestrians present significant real-world challenges.
• Current AV algorithms often lack sufficient consideration for human pedestrian preferences and safety perceptions.
• Developing methods to align AV behavior with human expectations is crucial for widespread adoption.

Purpose of the Study:

• To develop a methodology for directly eliciting and quantifying pedestrian-preferred AV behaviors.
• To create a feedback loop for adapting AV algorithms based on human judgments.
• To improve the performance and safety of AVs through human-centered design.

Main Methods:

• Utilized a Deep Q Network (DQN) in a simulated pedestrian crossing environment.
• Adapted the reward structure of the DQN using quantitative human behavioral feedback.
• Employed computational reinforcement learning (RL) and behavioral science techniques for iterative reward shaping.
• Collected data from 124 participants judging AV behaviors in a controlled setting.

Main Results:

• Demonstrated strong initial improvements in the assessment of AV behaviors using the adaptive reward structure.
• Identified that the predictability of AV movements is a key factor influencing positive human judgments.
• The iterative feedback loop effectively refined AV behaviors based on explicit and implicit human preferences.

Conclusions:

• Pedestrian judgments provide valuable data for enhancing AV behavior.
• Prioritizing predictable AV movements is essential for improving human-AV interactions.
• This methodology offers a scalable approach to developing safer and more acceptable automated vehicles.