Arkady Zgonnikov

What would a careful and competent human driver do? I build models that answer this question — to understand the human mind, and to hold automated vehicles to a human standard.

Assistant Professor · Department of Cognitive Robotics, TU Delft
Director for Partnerships · Centre for Meaningful Human Control
Co-director · HERALD Lab

Arkady Zgonnikov

Automated vehicles are already on our streets. The hard question is no longer whether they can drive — it is whether they drive well enough.

Regulations and industry practice ask automated vehicles to be at least as safe as a careful and competent human driver. That is an intuitive benchmark, but turning it into concrete, measurable requirements needs models that capture how people actually perceive, decide, and act in complex, dynamic, high-stakes situations. Driver models have been around since the 1930s, yet they still fall short of the richness and variability of real human driving.

I combine cognitive science with robotics and AI to build high-fidelity models of human driver behavior. These models do two jobs at once: they advance our fundamental understanding of the human mind, and they provide scientifically grounded safety references for those who develop and regulate automated vehicles. Throughout, I am guided by the principle of meaningful human control — increasingly capable autonomous systems should remain aligned with human intentions, norms, and values, including the most basic one: safety.

Three research lines

01
From cognitive science to driver behavior

How do drivers choose between competing actions — turn in front of an approaching car, or wait for it to pass? I established a new direction in driver behavior research by bringing in evidence accumulation models from cognitive neuroscience, together with new model architectures and experimental paradigms for measuring cognitive processes during driving. This work built a bridge between fundamental research on human decision-making and the study of driving.

02
Unified models of driving

Isolated decisions are only part of the story. By combining theories of human cognition with robotics techniques such as model-predictive control, I develop unified, closed-loop models that capture driving behavior across scenarios. With Nissan, this led to models grounded in risk homeostasis theory, progressing from simulator experiments to real-world traffic data. With Waymo, it led to a framework grounded in active inference, culminating in a collision-avoidance model with unprecedented generalization capabilities.

03
Meaningful human control over automated vehicles

Here I connect philosophical work on control and responsibility with hands-on empirical research on automated driving: developing methods for assessing whether — and to what extent — humans retain meaningful control over automated vehicles, and for quantifying responsibility in interactions between humans and AVs. This line provides the ethical foundation for the rest of my work.

Beyond the core. I also enjoy exploring adjacent areas: methods and open-source tools for evaluating trajectory prediction models and interaction-aware controllers, novel AV algorithms, and fundamental AI methods.

The next ten years

My aim is to build a science of human-referenced evaluation for autonomous systems, along three directions:

Deeper models of the driving mind

How should visual attention enter models of driving? What role does metacognition play? How far can one cognitive model generalize — and can machine learning make such models scale without losing their explanatory power?

From models to evaluation tools

Public-domain behavior reference models and AV evaluation methods that policymakers, certification authorities, and consumer testing organizations can use directly.

Ethical behavior references

Real driving involves trade-offs between safety, comfort, efficiency, and social norms. Models of how people navigate them can let us evaluate not only how safely an AV drives, but how it handles conflicts between human values.

I find it particularly rewarding to see fundamental research turned into tools that are used in the real world — so I have deliberately built partnerships across industry and the public sector to make that happen.

From cognitive modeling to deployed AVs
The Waymo Reference Driver

Developed in our collaboration with Waymo, the Reference Driver is a production-level model of how humans avoid collisions. Waymo uses it daily for virtual safety assessment of its automated vehicles — a direct pathway from cognitive modeling research to the safety assurance of deployed AVs.

Industry
Directly funded projects, joint proposals, and MSc projects with Waymo, Nissan, BMW, Porsche, TNO, and cogniBIT — including three patents in Japan and the US.
Public sector
Work on evaluating and regulating automated driving with the European Commission Joint Research Centre, the Dutch Vehicle Authority (RDW), and the Central Bureau for Driving Licenses (CBR).
Standards
Member of the SAE J3330 task force, developing recommended practice for behavior reference models in automated driving assessment.
Policy
Work on meaningful human control has been incorporated into EU and global policy documents on autonomous weapon systems and AI for crisis management.

Where this is heading

There is strong demand from manufacturers, regulators, and policymakers for scientifically grounded, ready-to-use tools for AV safety assessment. My ambition is to meet it by translating driver modeling research into public-domain behavior reference models and evaluation methods that can be widely adopted in industry practice, regulation, certification, and consumer testing.

Also in the media: Delta, Algemeen Dagblad, NEMO Kennislink, TU Delft.

In my ideal research team, people have the autonomy to develop their own ideas, the support to pursue them, and the opportunities to build careers beyond the group.

I see my role as providing direction and opportunities while giving team members ownership of their projects — and the freedom to experiment, take risks, and make mistakes. The group's intellectual direction is shaped by their questions and expertise, not imposed top-down. My team is part of the HERALD Lab, which I co-founded with Luca Laurenti, and works closely with Simeon Calvert's ADAS Lab to build a larger community around human-centered automated driving.

Where people go next

2
PhD alumni moving on to tenure-track positions (TU Delft, Sampoerna University)
4
MSc alumni continued to PhDs (TU Delft, Utrecht, Aalto, Chalmers)
→
Others now research driver behavior at Waymo, Jaguar Land Rover, RDW, and cogniBIT

PhD candidates

Ashwin George
Quantifying responsibility in spatial multi-agent interactions
Sahel Akbari (Erasmus MC)
Data-driven assessment of upper-limb movement after stroke

PhD alumni

Julian Schumann cum laude
Evaluating predictions and explaining human driving behavior
Lucas Elbert Suryana
Meaningful human control in automated vehicle decision-making
Olger Siebinga cum laude
Communication-enabled interactions in highway traffic

Postdocs & researchers

Bhargav Nallapu
Postdoc, 2024–2025
Federico Scari
Research engineer, 2024–2025
Samir Mohammad
Research engineer, 2023–2024

The thing I enjoy most about my work is interacting with students — their curiosity, creativity, and independence.

Students are exceptionally good at finding creative solutions to complex problems, but it is easy to become so immersed in solving a problem that the bigger picture disappears: why it matters, what gap it addresses, what follows from it. In research projects I coach students to find a real gap in knowledge and to recognize that they are not merely implementing a model or running an experiment, but possibly becoming the first person to discover something about the world. In lectures, I connect technical concepts to the human and societal consequences of robotics. And I encourage students to look beyond their own discipline, because that is where better questions come from.

The goal is not only to train people who can build sophisticated technology, but engineers and scientists who know which problems are worth solving — and why.
Graduation projects
Over 35 MSc projects supervised across TU Delft faculties, many leading to publications; one received the award for the best AI-related MSc thesis at TU Delft (2024)
Internship coordinator, MSc Robotics
Around 500 students coordinated to date
2021 – present
Fundamentals of AI
Responsible for the project component of the interdisciplinary, university-wide MSc programme
2021 – 2024
Human-Robot Interaction
Invited lecture on human behavior modeling
2021 – present
Robots and Society
Invited lecture on meaningful human control
2021 – present

I enjoy bringing people together around shared goals — across research groups, disciplines, and the boundary between academia and industry.

Building a new research community

Cognitive models for AV safety assessment are an emerging niche at the intersection of cognitive science, AI, and automated driving. I have been part of growing it from a small meeting of experts into a recurring international event:

2024
Invited to a two-day expert workshop at Wittenburg Castle to define the emerging research agenda
2025
Keynote at the follow-up workshop at IEEE IAVVC
2026
Main organizer of the workshop at IEEE ITSC — now a yearly event series, with me on its steering committee

Centre for Meaningful Human Control

In 2024 I co-founded the Centre for Meaningful Human Control with Luciano Siebert and David Abbink, bringing together around 30 researchers across engineering, design, ethics, and philosophy. As Director for Partnerships and Collaborations, I connect members to research opportunities and build partnerships with organizations that can put our research to use.

At TU Delft and beyond

Automated driving community at TU Delft
Seminars with invited experts including Missy Cummings, Cristina Olaverri-Monreal, and Johan Engström; events with Mondai House of AI and the TU Delft AI Initiative
IEEE Intelligent Transportation Systems Conference
Associate editor
2026
Reviewing
IEEE T-ITS and T-IV, Transportation Research, Human Factors, Psychological Science, Cognition, ACM/IEEE HRI, and others