Emotion in an active inference model of human driving

arXiv:2608.07480 · cs.AI, cs.HC, cs.LG, cs.RO · Submitted 2026-06-09 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Emotion in an active inference model of human driving".

Jane: The paper was written by Julian F. Schumann, Johan Engström, Ran Wei, Jens Kober, Martijn Wisse et al. from Delft University of Technology and Waymo LLC.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Jane: We also have Lu with us today — senior AI researcher at Tsinghua.

Tom: We also have Meng with us today — lead engineer at a mysterious AI startup.

Jane: We also have Lalam with us today — the in-house Large Language Model.

Tom: Alright, let's get started.

Title: Tom: Welcome back to the show, everyone! Today we are digging into a paper that just hit arXiv, and the title alone got me hooked: "Emotion in an active inference model of human driving."

Jane: Oh, Tom, that title is a mouthful, but it's such a cool mashup. You've got emotions, you've got driving, and you've got this thing called active inference. For our listeners, active inference is basically this theory that our brains are constantly trying to predict what's going to happen next and then acting to make those predictions come true.

Tom: Right! And this paper is trying to put feelings into that framework. I mean, we all know driving is emotional. Someone cuts you off, you get angry. You're cruising on an empty road, you feel relaxed. But modeling that in a computer is really hard.

Jane: Exactly. And the authors here—Schumann, Engström, Wei, Kober, Wisse, and Zgonnikov—they're from Delft University and Waymo. So you've got a great mix of academic research and real-world autonomous vehicle engineering.

Tom: Waymo being the self-driving car company, right? So this isn't just abstract theory. They want to understand how human drivers feel so they can maybe build better autonomous systems that interact with us.

Jane: That's the big implication for me. If a self-driving car can predict that a human driver is getting anxious or angry, it can change its own behavior to defuse the situation. That could make the roads safer for everyone.

Tom: And they're not just guessing. They're using this active inference model, which is all about how agents—people or robots—reduce uncertainty. The paper says emotions like valence and arousal can be pulled out of that model's internal states.

Jane: Valence is the positive or negative feeling, and arousal is how alert or activated you are. So anger is negative valence and high arousal, while calm is positive valence and low arousal.

Tom: So they're basically saying, "Hey, we can read the emotional state of a simulated driver just by looking at its predictions and uncertainties." That's a pretty bold claim.

Jane: It is, but they've got the math to back it up. And I'm really curious to see how they actually tested it. What scenarios did they use to make the simulated driver feel something?

Tom: I saw they mention lateral incursions and intersections with priority violations. Sounds like they put their virtual driver in some pretty stressful situations.

Jane: Which is exactly what we need to hear about. Let's dig into the summary and see what they found.

Summary: Tom: So, Jane, we've got the title. Let's talk about what the paper actually does. The summary in the abstract is pretty dense, but the core idea is that they extended an existing model of driving behavior to include emotions.

Jane: And the key move they made, Tom, is that they didn't just look at the current state of the driver. They looked at what the driver *predicts* will happen in the future. That's a big deal.

Tom: Right, because your feelings right now depend on what you think is coming. If you see a car swerving toward you, you're not just reacting to the swerve; you're predicting a crash. That's what drives the fear.

Jane: Exactly. So they define valence as the difference between the actual value of the situation and the expected value. If things are going better than you expected, you feel positive. If they're going worse, you feel negative.

Tom: And arousal, they tie to uncertainty. The more uncertain you are about whether a collision is going to happen, the more aroused you get. That makes total sense.

Jane: It does. And they test this in two scenarios. One is a lateral incursion, where another car suddenly veers into your lane. The other is an intersection where someone runs a stop sign or a yield sign.

Tom: And they compare what happens when the other driver is compliant—does what they're supposed to—versus when they're adversarial and break the rules.

Jane: The results are pretty intuitive, which is a good sign for the model. When the other driver is compliant, the simulated driver stays calm. But when the other driver violates norms, you see a spike in arousal and a drop in valence.

Tom: So the model gets "angry" when someone cuts it off. That's actually a really nice validation that the math is capturing something real.

Jane: And they even map these emotions onto the circumplex model, which is that classic circle with valence on one axis and arousal on the other. They show the simulated driver moving from "calm" to "angry" and then to "relaxed" after avoiding a crash.

Tom: That's the kind of result that makes you sit up and take notice. It's not just a number; it's a recognizable emotional journey.

Jane: And it's all happening inside a model that's already being used to study human driving behavior. So they're adding an affective layer on top of a cognitive layer.

Tom: I love that. But I'm also wondering about the details. How did they actually implement this? The summary mentions continuous state spaces and multiple time steps. That sounds complicated.

Jane: It is, but it's also necessary. Real driving isn't a simple grid of states. It's continuous. And their approach seems to handle that gracefully.

Tom: Let's get into the improvements they made over previous work. That's where the real meat is.

Improvements: Jane: Tom, we've established that this paper adds emotions to a driving model. But what's genuinely new here compared to earlier attempts at modeling emotion with active inference?

Tom: Well, the paper is very clear that previous work, like the one by Pattisapu and colleagues, was limited to simple, discrete state spaces. Think of it like a board game where you move from square to square.

Jane: And this paper moves it to continuous states, which is more like real driving where you have smooth positions and velocities. That's a huge step up in complexity.

Tom: Right. And they also had to deal with the fact that their driving model plans over a horizon of multiple future steps. The old emotion model only looked at the immediate next step.

Jane: So they had to figure out how to compute valence and arousal when the driver is thinking ten steps ahead. They did this by averaging the expected value over several past policies and comparing it to the actual value of the current situation.

Tom: And they were careful about arousal. They didn't just look at raw kinematic uncertainty, because that gives weird results. For example, a car far away is uncertain, but it shouldn't make you aroused because it's not a threat.

Jane: Right, that would be silly. Instead, they defined arousal based on the predicted probability of a collision. That's a much more meaningful abstract state.

Tom: So they're saying, "Let's not get distracted by all the noise in the raw sensor data. Let's focus on the thing that actually matters for survival: am I going to crash?"

Jane: Exactly. And that's a really smart design choice. It makes the emotion signal interpretable. You can look at the arousal and immediately understand why the driver is stressed.

Tom: They also mention making the valence calculation more stable by averaging over multiple past policies. That smooths out the randomness from the particle filter they use.

Jane: Which is important for a model that's supposed to be used in simulations. You don't want the emotions to be jumping around randomly frame to frame.

Tom: So the improvements are about making the emotion model work in a realistic, continuous, multi-step planning context. And they did it without breaking the underlying driving model.

Jane: And the results we saw in the summary suggest it works. The emotions track the scenario in a way that matches human intuition.

Tom: But I want to know more about the actual experiments. The first page of the paper sets up the problem. Let's look at that to see how they frame the whole thing.

First Page: Tom: Alright, Jane, let's go back to the very beginning of "Emotion in an active inference model of human driving." The first page sets the stage by talking about why active inference is such a big deal.

Jane: It does. They describe it as a first-principles account of adaptive behavior. That means it's not just a bunch of rules; it's a fundamental theory about how agents minimize surprise.

Tom: And they mention that active inference has been applied to everything from single cells to plants to human cognition. That's a pretty broad claim, but it shows the framework is versatile.

Jane: Then they zoom in on driving. They mention previous work on car-following, lane-keeping, collision avoidance, and intersection interactions. So this is an established line of research.

Tom: And the key gap they identify is that none of those models account for emotions. Which is wild, because driving is so emotional.

Jane: They also point out that existing emotion models in active inference were limited to toy problems. So they're positioning their work as the bridge between theoretical emotion modeling and realistic driving scenarios.

Tom: And they mention the practical applications. This isn't just for fun. They talk about using these models in simulation-based testing for autonomous vehicles.

Jane: That's a big deal. If you're testing a self-driving car, you want to make sure it can handle an angry human driver, not just a rational one.

Tom: The first page also introduces the circumplex model of affect, which we've talked about. And they mention that it's been used in driving contexts before.

Jane: So they're building on a solid foundation. They're not inventing a new theory of emotion; they're taking a well-known one and fitting it into the active inference framework.

Tom: And they're doing it in a way that's continuous and multi-step, which is what makes it applicable to real driving.

Jane: I also noticed they mention the work by Pattisapu et al. as their starting point. So they're being very clear about their intellectual lineage.

Tom: That's good science. You stand on the shoulders of giants, but you also have to point out where the giants stopped.

Jane: And they stopped at discrete states. This paper takes the next step.

Tom: So the first page is really about setting up the problem: we have a great framework, we have a great emotion model, but they haven't been combined in a realistic way. That's what this paper does.

Jane: And the experiments we discussed show it works. The simulated driver gets angry, gets scared, and feels relief.

Tom: Let's wrap this up and give our final thoughts.

Conclusion: Tom: We've been talking about "Emotion in an active inference model of human driving" all episode, and I think it's fair to say this is a significant step forward.

Jane: Absolutely, Tom. The paper takes a well-established framework for modeling behavior and successfully layers emotions on top of it. They showed that valence and arousal can be derived from the model's internal predictions and uncertainties.

Tom: And they validated it in two realistic scenarios—a lane incursion and an intersection violation—where the simulated driver's emotions matched what we'd expect from a human.

Jane: The implications are pretty exciting. For autonomous vehicle development, this could mean cars that understand when a human driver is stressed or angry and adapt their behavior accordingly.

Tom: And for psychology, it gives us a mechanistic account of how emotions arise from cognitive processes. That's a big deal for understanding human behavior.

Jane: There are limitations, of course. They didn't compare their model to actual human emotional responses, and they didn't model how emotions feed back into driving behavior. But that's future work.

Tom: Right, they even mention that angry drivers might take more risks, and that could be a next step. But for now, this is a solid foundation.

Jane: So, we say goodbye to this paper. It's a great read for anyone interested in the intersection of cognitive science, robotics, and emotion.

Tom: And we're looking forward to seeing where this line of research goes. Thanks for listening, everyone. We'll be back with another paper soon.

Jane: See you then!

Julian F. Schumann, Johan Engström, Ran Wei, Jens Kober, Martijn Wisse, Arkady Zgonnikov

Delft University of Technology · Waymo LLC

cs.AI, cs.HC, cs.LG, cs.RO

Submitted: 2026-06-09

Updated: 2026-08-11

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 59/100

Key concepts

Active Inference
A theory suggesting that brains constantly predict future events (minimize surprise) and then act to make those predictions come true. It serves as the fundamental framework for modeling adaptive behavior.
Valence
A feeling described as the positive or negative quality of a situation. It is defined in the model as the difference between a situation's actual value and its expected value.
Arousal
A measure of how alert or activated a person is. In this model, it is tied to uncertainty, specifically based on the predicted probability of a collision.
Circumplex Model of Affect
A classic representation used to map emotions using two axes: valence (positive/negative) and arousal (alert/calm). The model shows simulated drivers moving through recognizable emotional states.

Terminology

Summary

Summary

This paper extends an existing active inference model of human driving to incorporate a model of driver emotions, addressing a gap in prior work where affective states were not accounted for in such models. The authors note that an important aspect of human driving behavior – so far unaccounted for in these active inference models – is the role of emotions, and that previous work suggests that human behavior and driving capability is highly influenced by affective states. While prior active inference-based emotion models exist, they are limited to simplified toy problems with discrete state spaces and lack the capacity for realistic planning, which is generally needed to model human road user behavior in most traffic scenarios.

The paper builds on the circumplex model of affect by Russell, which places emotions along the dimensions of valence and arousal. The authors extend the work of Pattisapu et al., who mapped active inference quantities onto these dimensions, generalizing their formulation to a continuous state space and accounting for policies of actions planned over multiple time steps. In the original formulation, valence corresponds to the difference between expected and realized utility, while arousal is given by the entropy of the agent's posterior belief.

The underlying active inference model of driving, based on prior work by Schumann et al., uses a particle filter to represent beliefs, with a generative model that includes a kinematic likelihood based on the bicycle model and a normative likelihood biasing the model to assume other agents comply with traffic norms. The agent selects actions by evaluating candidate policies based on expected free energy, which balances pragmatic value (achieving goals) and epistemic value (reducing uncertainty). The agent updates its policy only when accumulated surprise against the current policy exceeds a threshold.

The proposed emotion model introduces several modifications. For valence, the authors divide the calculation by the time step to ensure independence from time step frequency, and they incorporate a prediction horizon of H > 1, consistent with appraisal theory which posits that emotions can be influenced by assessments of future events. The expected value is computed over the previous N = 5 policies, using a common prediction horizon shared by all considered policies. The actual value is calculated based on the reference policy currently in effect, preventing unimplemented counterfactual policies from inducing affective changes. The new valence term is defined as the difference between actual and expected value.

For arousal, the authors reject a naive generalization of the original formulation that would link kinematic uncertainty directly to arousal, as this would produce counterintuitive results (e.g., a nearby agent on a collision course would evoke lower arousal than a distant agent). Instead, they compute arousal over meaningful abstract states, specifically the predicted likelihood of collision. The arousal is defined as the binary entropy function of the collision probability, with the collision probability being the maximum over the predicted future states of a binary collision classifier based on the separating axis theorem.

The proposed approach is evaluated in two interactive driving scenarios, each run 128 times with half involving a compliant other agent and half involving an adversarial other agent. In the lateral incursion scenario, the modeled agent drives at 40 mph on a road with opposing lanes, and the other agent either stays in its lane or executes a lateral incursion maneuver. In the intersection scenario, the modeled agent approaches an orthogonal intersection with priority, and the other agent either brakes compliantly or violates the traffic rules.

In the lateral incursion scenario with a compliant other agent, the results show an initial increase in arousal as the agents approach, peaking around t = 1.4 s, followed by a decrease as the predicted likelihood of collision diminishes. This leads to positive valence peaking around t = 4.2 s. In the adversarial setting, arousal increases consistently after the other agent begins its incursion, and valence becomes increasingly negative. The agent changes its policy to avoid the collision, with arousal peaking at t = 3.8 s before decreasing as the agent becomes more certain of avoiding the collision. The agent reconsiders its policy at t = 4.6 s, and the second avoidance attempt results in an immediate drop in arousal.

In the intersection scenario with a compliant other agent, arousal and collision contribution to pragmatic value remain negligible, and valence remains neutral. In the adversarial case, the agent starts to lose trust in the other agent's norm compliance after t = 3 s, reflected in increasing arousal and negative valence. At t = 5 s, the agent abandons its priority and brakes, leading to a drop in arousal and an increase in valence. In collision outcomes, the decrease in arousal is caused by certainty that a collision has become unavoidable, and valence starts to increase later and reaches a lower peak than in the successful avoidance case.

Mapping the emotions onto the circumplex model, the authors find that in the lateral incursion scenario, emotions remain stable when the other agent is compliant, but exhibit large fluctuations in the non-compliant case, with the agent initially responding with anger before returning to a relaxed state after avoiding the collision. In the intersection scenario, the agent remains calm in the benign case, but in the adversarial case reacts with anger, shifting to sadness following the braking maneuver before returning to a relaxed state. In the collision scenario, the emotional state transitions from sadness to depression.

The authors discuss that the resulting affective responses are consistent with intuitive expectations, noting that non-compliant, norm-violating – and therefore unexpected – behavior by the other agent induces anger (i.e., negative valence and increased arousal), which is consistent with previous empirical studies. The primary exception is the decrease in arousal immediately prior to the collision, which the authors explain by noting the model "accounts only for the resolved uncertainty over the high-level outcome (i.e., whether a collision occurs), it neglects the persistent uncertainty about the consequences of that collision (e.g., injury or death) that might cause elevated arousal in reality at this stage."

The authors highlight that their method is consistent with appraisal theories of emotion, which emphasize that affective responses are driven by anticipated outcomes and the role of social and normative expectations. They also note that the approach is applicable to a general class of active inference models with discrete time and continuous state representations, and that incorporating information from multiple past policies reduces variability in affective signals.

The paper acknowledges several limitations. First, the model captures valence as a function of short-term prediction error, neglecting the influence of higher-level or long-term expectations, such as sustained discrepancies relative to prior expectations. Second, the work is limited to simulation and has not been compared to human data. Third, the model focuses on the effect of behavior on emotion, while potential effects of emotion on behavior are out of scope, though the authors suggest this could be addressed by modulating preference priors under negative valence.

Improvements for AI systems

Based on the paper, here are the specific improvements I can implement in AI systems:

Improvement: Integrate the valence-arousal emotion model into the active inference framework of AV decision-making.

What the improved system can do:

  • Continuously estimate the emotional state of human drivers (both the AV's own driver and surrounding human drivers) using only kinematic observations (position, velocity, acceleration)

  • Detect norm violations by other road users (e.g., running red lights, unsafe lane changes) and trigger appropriate emotional responses (anger, increased arousal) that modulate risk assessment

  • Predict when a human driver is likely to experience anger or stress based on unexpected behavior of other agents, allowing the AV to adjust its driving style (e.g., give more space, yield earlier) to de-escalate potentially aggressive situations

These improvements are directly implementable in existing active inference frameworks and can be validated against the two scenarios described in the paper (lateral incursion and intersection priority violation).

Abstract

Active inference has emerged as a principled framework for modeling adaptive behavior by balancing goal-directed action with uncertainty reduction. It has been successfully applied across biological and artificial systems, including recent work on human driving. However, existing active inference models of driving have yet to address an important determinant of behavior in traffic: affective state, which significantly influences decision-making. Prior work in non-traffic domains has explored active inference agents in which emotions are represented along the axes of valence and arousal in the circumplex model. However, this work has been limited to simplified settings with discrete state spaces. In this work, we propose an expanded formulation of valence and arousal that can be extracted from a more complex active inference model of driving with continuous states. In particular, we condition affective estimates not only on the current state but also on predicted future outcomes. We evaluate the proposed approach in two interactive driving scenarios and show that the resulting emotion signals correspond to affective patterns reported in similar scenarios.

Sources

Related papers