Pascal Jansen

Research

I work toward calibrated human-AI interaction.

Embodied AI, such as automated vehicles and service robots, increasingly acts in everyday life. These systems could make roads safer, give mobility to people who cannot drive, and take over tedious work. However, their effectiveness, people’s acceptance, and safe use often depend on human experience, such as trust, mental workload, and perceived safety.

These experiences can have an appropriate range. This range may depend on the person, the AI, and the situation. For instance, human experience can be inappropriate to the AI’s actual capabilities, as in undertrust or overtrust. The AI’s behavior can also be inappropriate to the person’s preferences, as in a driving style that feels too aggressive or a service robot that approaches too closely.

Interfaces that mediate (add, remove, or modify) information might calibrate human-AI interaction. However, people differ, so one design typically cannot serve everyone. In addition, people and AI change. A person may learn how the AI behaves, and the AI’s capabilities may change with each software update.

People in a public space seeing a service robot through different mediated-reality views.

To calibrate despite these obstacles, user interfaces should follow two principles:

Homeostatic adaptation

In physiology, homeostasis is the process by which organisms maintain vital conditions within a stable range despite a changing environment.1

1 Davies, Adaptive Homeostasis, Molecular Aspects of Medicine, 2016.

high low time → experience appropriate range AI or situation changes interface mediates experience stays within its appropriate range AI or situation changes, experience drifts interface mediates, experience returns

Human experience is the vital condition, and the AI and the situation are the changing environment. Ratings, physiological signals, and behavior can indicate when experience leaves its appropriate range. The interface can then mediate information until experience returns.

Hormetic doses

In toxicology, hormesis describes how controlled exposure to a stressor can stimulate adaptation, while excessive exposure can cause harm.2

2 Calabrese & Baldwin, Defining Hormesis, Hum. Exp. Toxicol., 2002.

high low time → experience, too high (e.g., overtrust) appropriate range harm small dose small dose small dose one large dose experience after one large dose interface reveals AI limit, a small dose more small doses, experience adapts too large a dose at once could harm

The AI’s limits are the stressor, and the interface controls the dose. Revealing one limit at a time can move experience toward the AI’s actual capabilities, for example from overtrust to calibrated trust. Revealing all limits at once could cause harm.

? ? ? Human passenger, resident, ... trust mental workload perceived safety ... Overtrust Overload , ... Adaptive interface display, voice, ... Interface display, voice, ... Embodied AI automated vehicle, cleaning robot, ... system state, limits, intent mediated information add, remove, modify one size fits all feedback constraints on behavior commands ? ? ?

I develop this vision through empirical and technical research across three areas:

Research areas around calibrated human-AI interaction Three overlapping circles. Embodied AI and Interfaces: extended reality, automated vehicles, urban air mobility, human-automation communication, human-robot interaction. Human Factors: trust calibration, situation awareness and uncertainty, accessibility and inclusion. Computational Design and Simulation: human-in-the-loop design optimization, simulation of user behavior, novel simulators. Overlaps: context-adaptive mediation, experience-guided optimization, scalable interface evaluation. Embodied AI& Interfaces HumanFactors ComputationalDesign & Simulation MIRAGE: Enabling Real-Time Automotive Mediated Reality (PDF)Extended Reality A Design Space for Human Sensor and Actuator Focused In-Vehicle Interaction (PDF)Automated Vehicles Fly Away: Motion Fidelity and Optimized Interface Design in Urban Air Mobility (PDF)Urban Air Mobility Improving External Communication of Automated Vehicles Using Bayesian Optimization (PDF)Human-AutomationCommunication Assessing Pedestrian Behavior Around Autonomous Cleaning Robots in Public Spaces (PDF)Human-RobotInteraction Longitudinal Effects of Visualizing Uncertainty of Automated Vehicles (PDF)Trust Calibration Visualizing Imperfect Situation Detection and Prediction in Automated Vehicles (PDF)Situation Awareness &Uncertainty eHMI for All: External Communication for Pedestrians, Drivers, and Cyclists (PDF)Accessibility &Inclusion OptiCarVis: Improving AV Functionality Visualizations Using Bayesian Optimization (PDF)Human-in-the-Loop Design Optimization PedSUMO: Simulacra of Automated Vehicle-Pedestrian Interaction (PDF)Simulation of User Behavior SwiVR-Car-Seat: Vehicle Motion Effects in Virtual Reality Automated Driving (PDF)Novel Simulators CalibratedHuman-AIInteraction Context-Adaptive Mediation Scalable InterfaceEvaluation Experience-GuidedOptimization

Embodied AI & Interfaces

Human Factors

Computational Design & Simulation

Where the Areas Meet

  • Context-Adaptive MediationEmbodied AI & Interfaces with Human Factors
  • Experience-Guided OptimizationHuman Factors with Computational Design & Simulation
  • Scalable Interface EvaluationEmbodied AI & Interfaces with Computational Design & Simulation

Featured Work

Toward homeostatic adaptation OptiCarVis is a first step toward homeostatic adaptation. It optimizes visualizations of what an automated vehicle detects, predicts, and plans. The person’s ratings served as feedback, and an optimizer proposed each next design from them. The optimized designs improved trust, acceptance, and perceived safety without an increase in mental workload.

OptiCarVis: Improving Automated Vehicle Functionality Visualizations Using Bayesian Optimization to Enhance User Experience

Pascal Jansen, Mark Colley, Svenja Krauß, Daniel Hirschle, and Enrico Rukzio. 2025. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery. https://doi.org/10.1145/3706598.3713514 Honorable Mention Award.

Toward hormetic doses This study is a first step toward hormetic doses. Its visualizations show an automated vehicle’s uncertain detections and predictions in situ. Each shows one limit of the AI at the moment it matters, a small dose. The 50 participants saw the visualizations twice a day for three days, and their trust and perceived safety increased over the three days.

Longitudinal Effects of Visualizing Uncertainty of Situation Detection and Prediction of Automated Vehicles on User Perceptions

Pascal Jansen, Mark Colley, Max Rädler, Jonas Schwedler, and Enrico Rukzio. 2025. Transportation Research Part F: Traffic Psychology and Behaviour. Elsevier. https://doi.org/10.1016/j.trf.2025.05.013

Toward the interface as mediator A mediating interface changes what the person perceives of the AI and its surroundings. MIRAGE realizes this idea for the view through the windshield. It adds, removes, and modifies objects in real time in real vehicles, built on models for object segmentation, depth estimation, and inpainting.

MIRAGE: Enabling Real-Time Automotive Mediated Reality

Pascal Jansen, Julian Britten, Mark Colley, Markus Sasalovici, and Enrico Rukzio. 2026. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery. https://doi.org/10.1145/3772318.3791195

Funding

The German Research Foundation (DFG) funds the research grant Computational Optimization of In-Vehicle User Interface Design for Calibrated Trust and Maximized Usability and Safety with EUR 388,213. The principal investigator is Prof. Dr. Enrico Rukzio. I contributed substantially to the conception and writing of the proposal. The project extends my doctoral research.

I co-led and implemented work in SituWare, a project on driver situation awareness in highly automated driving funded by the German Federal Ministry for Economic Affairs and Energy (BMWi).