Improving External Communication of Automated Vehicles Using Bayesian Optimization

Research Background and Issues

• Identified Problems or Challenges:
  Autonomous vehicles (AVs) lack the non-verbal communication methods of human drivers (e.g., gestures, eye contact), leading to uncertainty during interactions with other road users, such as pedestrians. Addressing these issues requires external human-machine interaction interfaces (eHMIs). Although various eHMI design solutions have been explored, such as LED strips, windshield displays, and auditory cues, there remains a lack of comprehensive understanding regarding how to optimize across multiple design parameters and user goals.

• Significance:
  The widespread adoption of AVs in future transportation systems depends on effective communication with other road users. Properly designed eHMIs can enhance pedestrians' trust, perceived safety, and understanding of AV intentions, whereas poor designs may lead to information misinterpretation or traffic accidents.

• Research Motivation and Related Work:
  While existing studies have demonstrated the positive impact of eHMIs on trust and safety perception, they often focus on single parameters or a limited number of design options. The optimization of multidimensional eHMI design parameters (e.g., color, size, position, sound intensity) to meet diverse user needs remains an open question. Additionally, individual differences, such as gender, and their influence on design requirements have been underexplored.

Solution

• Proposed Solution:
  The authors employed a "Human-in-the-Loop Optimization" approach based on multi-objective Bayesian optimization (MOBO) to explore and optimize eHMI designs. Their goal was to optimize multiple visual display parameters and auditory components to improve trust, safety perception, predictability, acceptance, and aesthetics while reducing cognitive load and pedestrian waiting time.

• Innovations:

  1. Multi-objective Optimization: Unlike traditional methods that optimize for a single design goal, MOBO enables simultaneous optimization of multiple objectives and generates optimal trade-off points (Pareto front).
  2. Gender Difference Analysis: Investigates whether eHMI design parameters vary significantly based on gender.
  3. Human-in-the-Loop Optimization: Incorporates user feedback into the optimization process, iteratively adjusting design parameters to enhance user satisfaction and adaptability.

• Implementation Steps:

  1. Developed a Unity-based virtual reality (VR) experimental environment to simulate pedestrian crossing scenarios on a two-lane road.
  2. Defined nine optimization parameters (e.g., color RGB, transparency, flashing frequency, component size and position) and seven optimization objectives (e.g., trust, safety perception, aesthetics).
  3. Used Bayesian optimization algorithms (BoTorch library) to optimize eHMI designs over 20 experimental iterations.
  4. Collected subjective quantitative evaluation feedback from 37 participants and analyzed the data to determine the Pareto front.

Research Outcomes

• Specific Results:

  1. Feedback from all participants indicated that the optimization process improved multi-objective scores (e.g., trust and safety perception) while reducing cognitive load and time costs.
  2. Gender comparisons revealed no significant parameter differences, suggesting that certain design parameters may be suitable for a broader audience.
  3. The study recommended "partially universal" design parameters as a starting point (e.g., 3Hz flashing, cyan tones, large frontal display area) with targeted optimization based on these foundations.
  4. Bayesian optimization demonstrated effectiveness in rapidly identifying design trade-offs in multi-objective optimization.

• Comparison with Existing Solutions:

  1. Unlike traditional eHMI design methods, this study employed a more systematic and data-driven optimization approach, offering more comprehensive and personalized design solutions.
  2. Compared to existing studies relying on single modalities (e.g., LED strips), this study showcased the advantages of multimodal designs by integrating visual and auditory signals.

• Experimental or Evaluation Results:

  1. Scores for objectives such as trust, safety perception, and acceptance significantly improved with optimization iterations.
  2. The linear decline in time-to-completion trends indicated that optimization reduced pedestrian task completion times.
  3. Open-ended feedback highlighted that users found eHMI sound design and multimodal designs particularly effective, though improvements are needed in transparency, gradual cues, and scenario diversity.

• Limitations and Future Directions:

  1. Participant Sample Limitations: Participants were primarily young students, and the applicability of the findings to other age groups and cultural contexts needs further validation.
  2. Environment and Context: Experiments were conducted solely in VR, and real-world conditions (e.g., lighting, weather) may affect eHMI effectiveness differently.
  3. Goal and Parameter Selection: The study revealed high correlations between certain goals (e.g., aesthetics and acceptance), suggesting future research could reduce redundant goals and focus on key factors.
  4. Personalization Improvements: Although partially universal parameters were identified, further exploration is needed to enhance adaptability for a wider range of users.

Conclusion

This study provides an innovative methodology for effective interaction between autonomous vehicles and pedestrians by employing multi-objective optimization to make eHMI designs more appealing, inclusive, and practical. The research emphasizes the importance of multimodal communication and offers practical guidance for future standardization and personalization of eHMI designs.