Linking Behaviour and Perception to Evaluate Meaningful Human Control over Partially Automated Driving

1. Problem Statement

Partial driving automation (SAE Levels 2–3) creates a fundamental tension: drivers remain legally responsible for vehicle behavior, yet their active control is significantly reduced. This reduction often leads to driver complacency (reduced vigilance) and a diminished Sense of Agency (SoA), undermining the driver's ability to intervene safely when automation fails.

While Meaningful Human Control (MHC) has been proposed as a normative framework to resolve this tension (requiring that automation tracks human reasons and that responsibility can be traced to a human), there is a lack of empirical methods to evaluate whether existing systems actually provide MHC. Current evaluations rely on post-crash analysis or generic user interviews, which fail to capture real-time interaction nuances or directly link objective behavioral data to subjective perceptions of control and responsibility.

2. Methodology

The authors conducted a controlled driving simulator study involving 24 participants to evaluate MHC under two distinct interaction strategies:

• Haptic Shared Control (HSC): Both the driver and automation apply torques to the steering wheel simultaneously, allowing for continuous collaboration.
• Traded Control (TC): Control authority is explicitly handed over; the automation takes full control unless the driver applies a torque above a specific threshold, at which point authority reverts to the driver.

Experimental Setup:

• Scenario: Participants drove a rural road with automated lateral control (steering only; speed fixed at 50 km/h). They performed overtaking maneuvers.
• Critical Event: A "silent automation failure" was simulated where the automated vehicle followed a trajectory leading to a collision with a motorcyclist. Participants were instructed to intervene when the system behaved unsafely.
• Data Collection:
  • Objective Behavioral Metrics: Derived from telemetry (e.g., reaction time, steering torque conflict, maximum torque, steering reversal rate, trajectory jerk, time-to-collision, number of takeovers).
  • Subjective Perception: Measured via post-trial Likert-scale surveys (1–10) based on four MHC properties (Shared Representations, Sufficient Control, Responsibility, Moral ODD) and post-experiment open-ended questionnaires.

Analysis Approach:
The study employed a mixed-methods approach:

1. Quantitative: Linear Mixed-Effect Models (LMMs) were used to test hypotheses regarding the relationship between control modes, behavioral metrics, and subjective scores.
2. Qualitative: Thematic analysis of open-ended responses to identify factors influencing MHC perception.

3. Key Contributions

• Methodological Framework: The paper proposes and validates a novel methodology for evaluating MHC by integrating objective behavioral metrics with subjective perception scores derived from established MHC properties. This moves MHC evaluation from abstract theory to empirical operationalization.
• Comparative Analysis: It provides the first direct comparison of HSC and TC regarding their ability to foster MHC, linking specific interaction dynamics (e.g., torque conflict) to driver perceptions.
• Design Guidelines: It offers concrete, evidence-based design recommendations for partially automated systems to enhance human control, responsibility, and safety.

4. Key Results

Quantitative Findings

• Control Mode Differences:
  • Reaction Time: Significantly shorter in HSC than in TC ($p < 0.001$), indicating higher engagement and readiness to intervene.
  • Takeovers: Significantly higher in HSC than in TC, suggesting more frequent active engagement.
  • Maximum Steering Torque: Significantly lower in HSC, indicating less physical effort and resistance required to override the system.
  • Subjective Scores: HSC yielded significantly higher scores for "Sufficient Control," "I understood AV," "Working together," "Safe," and "Trust" compared to TC.
• Correlations (Confirmatory Analysis):
  • Conflict vs. Understanding: A significant negative correlation was found between steering torque conflict and the perception that "the AV understood me." High conflict implies a misalignment of intent.
  • Reaction Time vs. Control (Surprising Finding): Contrary to the hypothesis that faster reactions indicate better control, a positive correlation was found between reaction time and the perception of "sufficient control." The authors suggest this may be due to the "sense of agency" being stronger when drivers make larger, later corrections compared to small, early adjustments.
• Non-Significant Findings: No significant differences were found between HSC and TC for trajectory deviation or jerk, and no strong correlations were found between behavioral metrics and the perception of "Responsibility."

Qualitative Findings

Three categories of factors were identified as influencing MHC perception:

1. Attitudinal Factors: Need for continuous vigilance (negative) vs. feeling of shared direction (positive).
2. Interaction Factors:
   • Positive: Subtle haptic guidance, continuous access to control, and effortless overrides (HSC).
   • Negative: Invasive forces, resistance against driver intent, and abrupt handovers (TC).
3. Trajectory Factors: Mismatched timing of maneuvers and unsafe trajectory deviations reduced the sense of "working together."

5. Significance and Implications

• Design Implications:
  • Prioritize HSC: HSC appears to support MHC better than TC by maintaining driver engagement and reducing the "vigilance decrement."
  • Effortless Intervention: Systems should be designed to allow effortless, smooth interventions. High resistance or "fighting" the automation erodes the sense of control.
  • Transparent Intent: Automation must communicate its intent clearly (e.g., through subtle torque feedback) to align with driver expectations.
• Responsibility Attribution: The study highlights a paradox where drivers feel responsible even when disengaged (TC), or feel forced into responsibility by system dominance (HSC). Designers must balance authority to ensure drivers feel accountable only when they have the capacity to act.
• Broader Impact: The proposed evaluation framework (Table 5 in the paper) can be applied beyond driving to aviation, robotics, and healthcare, providing a structured way to assess MHC in any human-automation system.

Conclusion:
The study demonstrates that Meaningful Human Control is not just a theoretical concept but a measurable state influenced by specific interaction designs. Haptic Shared Control (HSC) outperforms Traded Control (TC) in fostering the perception of control, understanding, and collaboration, primarily by reducing conflict and maintaining continuous engagement. Future designs must prioritize transparent communication and effortless intervention to ensure safety and ethical accountability in partially automated driving.