PaperlayerView exact revision on arXiv

Source-linked AI summary

A Human-Like Pedestrian Model for Automated Driving Simulations

Ruofeng Wang, Patrick Ebel, Philipp Wintersberger, Antti Oulasvirta

arXiv:2609.29175v1cs.HC

TL;DR

Pedestrian models for automated-driving simulation rarely combine broad behavioral coverage, transfer across traffic contexts, and closed-loop deployment, despite their role in AV development and evaluation. This paper trains a theory-grounded pedestrian policy as a partially observable sequential decision process under human constraints. The model reproduces documented crossing behaviors, yields fewer collisions than CARLA’s default AI walker, and retains realistic performance in unseen environments while allowing fine-tuning.

  • Problem

    Few pedestrian models combine broad empirical behavior coverage, generalization across traffic contexts, and closed-loop simulator deployment, although such models support automated-vehicle safety evaluation.

    Source: §I · p. 1, §V · p. 11

  • Method

    The paper trains a theory-grounded pedestrian policy for closed-loop simulation, formulating crossing as sequential decision-making under perceptual, decision-interval, and motor constraints.

    Source: §VI · p. 12

  • Results

    The model reproduces ten documented crossing phenomena, produces fewer collisions than CARLA’s default AI walker, and retains realistic crossing performance in unseen environments with fine-tuning.

    Source: §VI · p. 12

  • Takeaways & Limitations

    The results support the model as a basis for simulator-ready pedestrian agents combining behavioral realism, closed-loop deployability, and adaptability for AV training, testing, and evaluation.

    Source: §VI · p. 12

    The study focuses on crosswalk crossing in CARLA and does not yet cover jaywalking, signal compliance, multi-pedestrian interaction, or richer social negotiation.

    Source: §V · p. 12

Abstract

from arXiv · show

Automated vehicles must be able to interact with pedestrians safely and efficiently across diverse traffic situations. Although driving simulators offer a scalable testbed for learning such capabilities, existing theory-inspired pedestrian models are narrow in scope and limited to go/no-go crossing decisions in single-lane settings. While data-driven approaches can predict pedestrian behavior in complex situations, they lack sufficient observations in rare, safety-critical scenarios. Here, we propose an approach to training pedestrian models in simulators so that learned policies generate demonstrably human-like behavior in realistic, complex traffic scenarios, including multiple lanes, heavy traffic, and dangerous driving styles. Our technical contribution is a novel definition of pedestrian-vehicle interaction as a partially observable Markov decision process (POMDP) with theory-grounded perceptual, cognitive, and motor constraints. It accounts for the highly adaptive nature of human behavior in traffic and simulates how people adjust their responses according to perceived danger, time pressure, and the complexity of the situation. When trained via deep reinforcement learning (RL) with domain randomization in a simulator, the model reproduces the broadest range of empirical findings shown so far on human crossing behavior, including gap acceptance, yielding acceptance, hesitation, and evasive speed adjustment. We show that learned policies transfer to unseen traffic environments, and can be further adapted to local traffic norms with finetuning. Together, these results establish a blueprint for simulator-ready pedestrian models that can support the development and evaluation of automated driving systems.

I. INTRODUCTION

The introduction identifies a need for pedestrian models that combine broad behavioral realism, generalization, and closed-loop simulator deployment. It presents a theory-grounded model that learns adaptive crossing behavior under human constraints for automated-driving simulation.

I. INTRODUCTION

  • Existing approaches include explanatory models, data-driven predictors, and simulation-oriented controllers, but these address different aspects of pedestrian behavior.The introduction notes that binary crossing models focus on go/no-go decisions, while prediction and simulation approaches have other primary purposes.Source: §I · p. 1
  • Few existing models combine broad empirical behavioral coverage with generalization across traffic contexts and closed-loop simulator deployment.The paper frames this combination as an open problem for simulator-ready pedestrian models.Source: §I · p. 1
  • The proposed model reproduces human-like crossing behaviors, from hesitation and gap acceptance to evasive actions and traffic-dependent crossing strategies.The introduction also describes adaptation to different traffic norms and improved alignment after fine-tuning.Source: §I · p. 1, Figure 1 · p. 2
  • The paper formulates pedestrian crossing as a POMDP using bounded observations, belief estimates, and actions that can change throughout the crossing episode.Its reward balances safety, crossing efficiency, waiting cost, and motor effort, allowing behavior to emerge without explicit handcrafting.Source: §I · p. 2
  • The model is intended for high-fidelity traffic simulation, where it can navigate to a destination while interacting with surrounding vehicles.The authors evaluate behavioral variation, safety outcomes, transfer to unseen maps, and fine-tuning to different traffic environments and norms.Source: §I · p. 3

II. RELATED WORK

This section organizes pedestrian-crossing research around empirical behaviors observed at different stages of a crossing. At initiation, those behaviors include hesitation, impatience, and context-sensitive acceptance of gaps and yielding cues.

II. RELATED WORK

  • The review organizes ten representative crossing phenomena into initiation, adaptation during crossing, and higher-level strategy selection.It compares existing models against those phenomena and summarizes behavioral coverage, generalization, and simulator suitability.Source: §II · p. 3, A. Empirical Crossing Phenomena · p. 3

1) Decisions to initiate crossing:

  • Longer waits tend to increase willingness to cross, while pedestrians may pause or slow down to appraise approaching vehicles before committing.The review describes these patterns as impatience and hesitation, respectively.Source: 1) Decisions to initiate crossing · p. 3
  • Gap acceptance generally rises with vehicle time to arrival; for a given time to arrival, some studies report higher acceptance for faster vehicles.In yielding scenarios, lower approach speed and a farther stopping position can signal yielding intent and promote crossing initiation.Source: 1) Decisions to initiate crossing · p. 3

2) Decisions during crossing:

The section describes adaptive responses after roadway entry, strategies for crossing multiple lanes, and the differing aims and evaluation limits of pedestrian-model families.

2) Decisions during crossing:

  • After entering the roadway, pedestrians may retreat from a tentative crossing or run forward or step back when danger arises.These responses address an unsafe gap or a potential collision during crossing.Source: 2) Decisions during crossing · p. 3

3) Crossing strategies:

  • Pedestrians tend to cross all lanes in one movement under light traffic but may use rolling crossing in denser or more complex conditions.Rolling crossing proceeds lane by lane, with possible waits in intermediate safe positions.Source: 3) Crossing strategies · p. 3

B. Pedestrian Crossing Models

  • The review groups pedestrian models as explanatory, predictive, or simulation-oriented according to their primary purpose.It compares which of ten crossing phenomena these approaches address.Source: B. Pedestrian Crossing Models · p. 4

1) Explanatory models:

  • Explanatory models estimate crossing decisions from traffic and pedestrian variables or represent mechanisms such as evidence accumulation, bounded rationality, and strategic interaction.Their implementations have mainly been assessed on crossing initiation or simplified interactions rather than continuous behavior in complex, multi-lane scenarios.Source: 1) Explanatory models · p. 4

2) Predictive models:

  • Predictive models learn from observational data to classify crossing intention or forecast motion, but are primarily evaluated on prediction accuracy rather than reproducing documented behaviors.Their performance can also decrease across dataset, location, or traffic-context shifts.Source: 2) Predictive models · p. 4

3) Simulation-oriented pedestrian models:

The section motivates a simulator-ready model that reproduces pedestrian behavior under human constraints. It describes a closed-loop policy whose perceptual, decision-timing, and motor limits shape actions in evolving traffic.

3) Simulation-oriented pedestrian models:

  • Simulation-oriented agents support scalable training and evaluation, but are generally not designed to reproduce documented crossing mechanisms under bounded human constraints.The proposed model targets closed-loop interaction in complex traffic while retaining theory-grounded decision mechanisms.Source: 3) Simulation-oriented pedestrian models · p. 4

A. Pedestrian Model

  • The pedestrian repeatedly perceives traffic, estimates vehicle motion, selects an action, and revises it as the scene evolves.CARLA maintains the full traffic state, while the pedestrian receives partial and noisy information and can stop, walk, run, or move backward.Source: A. Pedestrian Model · p. 4, Figure 2 · p. 5

1) Perceptual, cognitive, and motor constraints:

  • Perceptual limits restrict the amount and accuracy of vehicle information, and a 0.3 s decision interval represents intermittent human action selection.The model uses limited vehicle observations and Gaussian angular noise, then estimates vehicle states from noisy distance observations with a constant-acceleration Kalman filter.Source: A. Pedestrian Model · p. 4, 1) Perceptual, cognitive, and motor… · p. 5
  • Motor delay and action-dependent speed limits constrain how quickly decisions take effect and how pedestrians move.The modeled delay is stochastic, while target speeds differ for stopping, walking, running, and moving backward.Source: 1) Perceptual, cognitive, and motor… · p. 5

2) POMDP formulation:

The paper formulates pedestrian crossing as a bounded-optimal POMDP, using partial observations and a closed-loop action policy. Its rewards balance safe, timely crossing with human-like behavioral preferences.

2) POMDP formulation:

  • The latent state includes pedestrian and vehicle motion, plus crosswalk attributes such as lane count and length.The state includes vehicle positions, velocities, and accelerations, as well as pedestrian position and velocity.Source: 2) POMDP formulation · p. 5
  • The policy chooses among four locomotion commands using bounded observations rather than perfect global knowledge.Observations include noisy vehicle-motion estimates and task-related variables; pedestrian motion includes delayed actuation.Source: 2) POMDP formulation · pp. 5–6
  • Rewards combine safe, efficient crossing objectives with behavioral preferences and progress-based shaping.The task rewards successful crossing and penalizes collisions and time; further rewards penalize near-misses and costly actions.Source: 2) POMDP formulation · p. 6
  • Locomotion costs favor walking when safe while retaining running and retreating as available responses.Running and moving backward carry costs, while stopping and walking incur no extra locomotion cost.Source: 2) POMDP formulation · p. 6

1) CARLA environment:

The pedestrian policy is trained in CARLA with randomized crossings and traffic, then optimized with deep reinforcement learning. The evaluation protocol tests behavioral realism, closed-loop outcomes, and transfer or adaptation across environments.

1) CARLA environment:

  • CARLA training uses varied Town10HD road geometries, while weather, illumination, and traffic-light timing are held fixed.Background vehicles follow traffic signals, but pedestrian yielding is disabled; the policy does not observe traffic-light phase, and signal-response modeling is outside scope.Source: 1) CARLA environment · p. 6

2) Domain randomization:

  • Domain randomization varies crosswalk geometry, vehicle types and speeds, and evolving traffic conditions within the training map.Crosswalks are sampled per episode, and 35–45 vehicles have target speeds sampled from 20 to 50 km/h.Source: 2) Domain randomization · pp. 6–7

3) Reinforcement learning:

  • A model-free deep RL policy is trained with PPO and can be deployed from specified start and goal locations or further trained for target environments.The policy is trained to maximize expected discounted return; PPO uses a learning rate of 3 × 10−4 and discount factor γ = 0.99.Source: 3) Reinforcement learning · p. 7

C. Evaluation Protocol

  • Evaluation tests ten documented behaviors, closed-loop crossing outcomes against CARLA’s AI walker, and transfer or adaptation in new traffic environments.Transfer is tested on an unseen map, while norm adaptation uses lower speeds, lower density, and stronger vehicle yielding; outcomes and behavioral composition are compared.Source: C. Evaluation Protocol · p. 7

1) Behavioral evaluation:

Evaluation covers crossing decisions before roadway entry, behavior after entry, crossing strategy, closed-loop outcomes, and transfer or adaptation. Statistical analyses test density trends and summarize proportions and acceptance patterns.

1) Behavioral evaluation:

  • Behavioral realism is assessed across initiation, in-road adaptation, and crossing strategy, using human distributions or trends when available.The ten target behaviors include hesitation, impatience, gap and yielding acceptance, aborts, evasive movements, and one-stage versus rolling crossings.Source: 1) Behavioral evaluation · p. 7

2) Closed-loop outcome evaluation:

  • Closed-loop usability and safety are measured as successful crossing, no crossing, or collision, with CARLA’s AI walker evaluated under the same definitions.Outcome rates are calculated overall and by traffic density.Source: 2) Closed-loop outcome evaluation · p. 7

3) Transfer and adaptation evaluation:

  • Transfer and adaptation tests compare original-domain, zero-shot, and fine-tuned performance, including changes in crossing style.Tests use an unseen CARLA map and a modified environment with lower vehicle speeds, lower density, and stronger yielding.Source: 3) Transfer and adaptation evaluation · p. 8

D. Statistical Analysis

  • Binary outcomes are analyzed with logistic regression using ordered traffic density to test monotonic trends; proportion estimates use Wilson intervals where applicable.Acceptance curves are summarized from empirical distributions to characterize timing patterns rather than test a single binary contrast.Source: D. Statistical Analysis · p. 8

IV. RESULTS

Across ten target behaviors, the policy generally reproduces the direction of documented empirical patterns, with initiation behavior adapting to waiting time, vehicle dynamics, and yielding cues. In the reported tests, the model shows density-dependent hesitation and impatience, and matches human-data trends in gap and yielding acceptance.

A. Behavioral Evaluation

  • Across ten target behaviors, the policy generally changes in the same direction as documented empirical regularities, spanning initiation, in-road responses, and multi-lane strategy.Dynamic traffic evaluations covered 6,000 episodes, while controlled tests isolated specific traffic factors.Source: A. Behavioral Evaluation · p. 8

1) Crossing initiation behaviors:

  • Hesitation increased with traffic density, with odds of hesitation rising 11.7-fold per density level (95% CI [10.1, 13.7], p < 0.001).Hesitation was defined as pausing for at least one decision interval or briefly retreating before crossing.Source: 1) Crossing initiation behaviors · p. 8, Figure 3 · p. 9
  • The cumulative probability of accepting the same initially rejected gap increased with waiting time, consistent with impatience findings.The controlled test repeatedly presented the same initial TTA.Source: 1) Crossing initiation behaviors · p. 8, Figure 3 · p. 9
  • Gap acceptance increased with TTA, and both human data and model results showed higher acceptance at higher vehicle speed for comparable TTA.The model matched the qualitative trends despite differences in TTA ranges and crossing timing from the human-data setup.Source: 1) Crossing initiation behaviors · p. 9, Figure 4 · p. 9
  • The model reproduced earlier yielding-related crossing initiation for lower-speed vehicles at the same stopping distance and for earlier braking at the same speed.The stopping-distance comparison held vehicle speed high and contrasted stopping at 4 m versus 8 m.Source: 1) Crossing initiation behaviors · p. 9, Figure 4 · p. 9

2) Behaviors during crossing:

The model adapts its behavior during crossing as traffic grows denser or more dangerous, and shifts toward staged crossing in denser traffic. Closed-loop evaluation also compares crossing outcomes with CARLA’s AI walker and across traffic densities.

2) Behaviors during crossing:

  • Abort crossing becomes more frequent as traffic density increases (OR per level = 2.8, 95% CI [2.5, 3.0], p < 0.001).The trend is consistent with prior findings on retreat under denser traffic.Source: 2) Behaviors during crossing · p. 9, Figure 5 · p. 10
  • Back actions increase with traffic density (OR = 2.3, 95% CI [1.7, 3.2], p < 0.001), while run actions show a weaker trend (OR = 1.1, 95% CI [1.0, 1.2], p < 0.1).Run and back actions are also more frequent when minimum TTA is below 5 s, the defined potential-danger condition.Source: 2) Behaviors during crossing · pp. 9–10, Figure 5 · p. 10

3) Crossing strategies:

  • Rolling crossings become more frequent as traffic density increases (OR per density level = 2.3, 95% CI [2.2, 2.5], p < 0.001), implying fewer one-stage crossings.The pattern is consistent with prior findings that rolling crossing becomes more prevalent under high traffic density.Source: 3) Crossing strategies · p. 10, Figure 5 · p. 10

B. Closed-Loop Outcomes and Baseline Comparison

  • The proposed model achieves higher successful-crossing and lower no-crossing and collision rates than CARLA’s AI walker; denser traffic shifts outcomes toward lower success and higher no-crossing and collision rates.The evaluation counts successful crossings, decisions not to cross, and collisions, stratified by traffic density.Source: Figure 6 · p. 10, B. Closed-Loop Outcomes and Baseline… · p. 10

1) Overall and density-stratified outcomes:

The model has a lower collision rate than CARLA’s default AI walker and retains crossing competence on an unseen map. Fine-tuning improves map-transfer outcomes and shifts behavior in a lower-speed, lower-density, more yielding traffic regime.

1) Overall and density-stratified outcomes:

  • As traffic density increases, successful crossings decrease while no-crossing and collision rates increase; no-crossing odds rise by 15.2 per density level (95% CI [13.3, 17.3], p < 0.001).Collision odds also increase (OR = 1.4, 95% CI [1.3, 1.5], p < 0.001).Source: 1) Overall and density-stratified… · p. 10

2) Comparison with the AI walker:

C. Generalization and Adaptation

1) Transfer to an unseen map under similar traffic density:

  • On the unseen map, zero-shot success remains comparable at 82.1%, versus 82.3% on the original map; fine-tuning raises success to 86.3% and lowers no-crossing and collision rates to 5.5% and 8.2%.Zero-shot no-crossing and collision rates are 8.4% and 9.5%, respectively.Source: 1) Transfer to an unseen map under… · p. 11, Figure 7 · p. 11

2) Adaptation to different traffic norm and cultural settings:

  • After fine-tuning in the more yielding, lower-speed, lower-density regime, run and back actions become less frequent while walk actions increase, producing smoother, more continuous crossings.Hesitation, abort-crossing, and rolling-crossing rates also decrease.Source: 2) Adaptation to different traffic… · p. 11, Figure 7 · p. 11

V. DISCUSSION

The paper concludes that a theory-grounded, partially observable policy can generate adaptive crossing behavior and transfer to related simulated environments. It also identifies limits to the study’s scope and to claims about real-world generalization.

V. DISCUSSION

  • Hesitation, impatience, aborts, evasive speed adjustment, and density-dependent strategy selection emerge from partial observability, bounded perception, limited motor capabilities, and incentives balancing safety and efficiency.The action set includes stopping, walking, running, and moving backward, supporting adaptive crossing responses to changing traffic,,.Source: §V · p. 11
  • The results suggest a reusable base policy can be adapted to new environments, and transfer assessment should consider behavioral style as well as safety outcomes.In the more yielding target regime, adaptation appeared as smoother, more direct, and less evasive crossing behavior.Source: §V · p. 11
  • Closed-loop adaptive agents may support more realistic testing of whether automated systems yield appropriately, communicate clearly, and remain robust across road layouts and traffic conditions.The paper notes that fixed-path or scripted pedestrians may fail to expose hesitation, tentative entry, retreat, and evasive speed changes.Source: §V · pp. 11–12
  • The study is limited to crosswalk crossing in CARLA and does not demonstrate broad generalization to real-world pedestrian behavior across diverse traffic situations.The authors also note simplified perception, designer-specified rewards, and pattern-level rather than subject-level behavioral comparison.Source: §V · p. 12
  • Future work could extend coverage to signalized crossings, jaywalking, and multi-pedestrian interaction, and strengthen calibration with controlled experiments, immersive studies, or field data.The authors also propose component ablations and investigating variation in pedestrian characteristics and data-driven reward design.Source: §V · p. 12

VI. CONCLUSION

  • The model replicates ten empirically documented crossing phenomena and produces fewer collisions than CARLA’s default AI walker, while retaining performance in unseen simulated environments and supporting fine-tuning.The paper presents it as a basis for simulator-ready pedestrian agents for automated-vehicle training, testing, and evaluation.Source: §VI · p. 12
Loading 2609.29175v1…