Cognitively Inspired Pedestrian Model

• The model simulates pedestrian behavior by integrating hierarchical behavior trees and reactive social force mechanisms for realistic motion and decision-making.
• It incorporates noisy perception, bounded rationality, and reinforcement learning to mirror human cognitive constraints and risk assessment.
• Empirical validation demonstrates sub-2 m trajectory error and high decision accuracy, offering robust safety insights for autonomous vehicles and urban planning.

A cognitively inspired pedestrian model is a simulation or mathematical framework for pedestrian motion and decision-making that explicitly encodes principles, constraints, or algorithms drawn from human cognition. These models incorporate elements such as perception (with noise and limitations), bounded rationality, memory, sequential decision processes, and risk assessment, aiming to reproduce both qualitative and quantitative patterns of real pedestrian behavior across a wide spectrum of domains—from low-level path selection to high-level scene understanding and interaction with vehicles. Key advantages of these approaches include greater empirical realism and the ability to generate interpretable predictions or scenario variations for applications such as autonomous vehicle (AV) testing, urban planning, and safety analysis.

1. Hierarchical and Modular Architectures

Many state-of-the-art cognitively inspired pedestrian models employ a hierarchical structure that mirrors layered cognitive processes. A typical architecture consists of:

• High-level (Deliberative) Layer: Employs behavior trees (BTs) or finite-state machines to select discrete maneuvers based on world state, agent goals, intentions, and scenario context. Each step, the BT re-evaluates conditions and selects one maneuver (e.g., "Wait at Crosswalk," "Enter Crosswalk," "Increase Walking Speed") based on the agent's current situation and scenario logic. This supports modularity and human interpretability, and lets engineers script or override specific behaviors for scenario-based testing (Larter et al., 2022).
• Low-level (Reactive/Motion) Layer: Converts BT-selected maneuvers into motion plans, typically using an adapted Social Force Model (SFM) or optimization-based trajectory generation. Reactive forces encode collision avoidance, comfort zones, border repulsion, and scenario-specific logic. For example, the SFM computes a sum of adaptive driving force towards a waypoint and exponential repulsive terms from nearby agents and vehicles (Larter et al., 2022).

This architecture allows for both deliberative, scenario-dependent behaviors and real-time adjustment to dynamic environmental changes. It is particularly well-suited to simulation platforms such as GeoScenario Server, which require extensible and interpretable behavior modules.

2. Cognitive Mechanisms and Inspiration

Cognitively inspired models operationalize several key principles from cognitive science and behavioral psychology, such as:

• Behavior Trees and Structured Decision Logic: BTs offer modular, readable control flow, rapid reactivity to dynamic inputs, and clean separation between conditions and maneuvers. This avoids state-explosion in classical finite-state machines and opacity inherent in rule-based black-box approaches (Larter et al., 2022).
• Noisy Perception and Bounded Rationality: Sensory noise (e.g., angular or distance uncertainty in vehicle speed estimation) is modeled explicitly, often using Kalman filters or Bayesian updating, leading to belief states over the environment. Crossing decisions or path choices are optimized under these information constraints, yielding boundedly rational behavior (Wang et al., 2024, Wang et al., 2024).
• Hybrid Mechanistic–RL Approaches: Reinforcement learning is combined with mechanistic perception and cost/reward functions that encode cognitive constraints such as looming aversion, walking effort, time pressure, and loss/gain asymmetry (Prospect Theory-like) (Wang et al., 2024, Haque et al., 2024).
• Intermediate and Variable Goals: Human-like dynamic re-planning is captured by assigning short-range, adaptive waypoints ("variable goals") as stepping stones to the final destination, mimicking the human tendency to favor locally optimal detours and avoid abrupt last-minute reactions (Jain et al., 9 Jan 2025).
• Risk-based Utility Maximization: Models frequently encode the tradeoff between speed/progress and the risk of collision using formal expected utility objectives—often adapted from Prospect Theory to capture nonlinear risk attitudes and probability distortion (Haque et al., 2024).
• Cognitive Map and Incomplete Knowledge: Higher-order navigation (e.g., in buildings) is handled via a cognitive map—typically a partially observed graph of navigable spaces, subject to uncertainty and continuous updating as knowledge is acquired through exploration or perception events (Andresen et al., 2016, Chraibi et al., 2017).

3. Mathematical Formulations and Core Algorithms

Motion and Interaction Dynamics

The Social Force Model remains a common backbone for low-level motion, with modifications for cognitive realism:

$$m_i \frac{d \mathbf{v}_i}{dt} = m_i \frac{v^{0}_i \mathbf{e}^0_i - \mathbf{v}_i}{\tau_i} + \sum_{j \neq i} \mathbf{f}_{ij} + \sum_W \mathbf{f}_{iW} + \sum_{\text{veh}} \mathbf{f}_{i\text{veh}}$$

Here, the driving force is set by the chosen maneuver or intermediate goal, while repulsive forces are parameterized by calibrated constants to reflect personal space and vehicle avoidance (Larter et al., 2022).

Perception and Belief Modeling

Perceptual models range from direct line-of-sight computations (ray casting over angular sectors) for obstacle avoidance (Moussaid et al., 2011), to Bayesian filters aggregating noisy angular or distance cues for future vehicle motion estimation (Wang et al., 2024, Wang et al., 2024).

Reinforcement Learning with Cognitive Constraints

Advanced models encode the full decision process as a POMDP or MDP, with states including position, velocities, and uncertain estimates of other agents' motion. The action set may be discrete ("Wait"/"Go") or continuous (desired crossing speed), and rewards combine positive utility for safe progress, negative utility for collisions, penalties for actuation effort, and cognitive or time-pressure terms. PPO or Double-DQN architectures are commonly employed (Wang et al., 2024, Wang et al., 2024, Trinh et al., 2019).

Utility-Based and Prospect-Theoretic Models

At each step, the agent chooses speed and heading to maximize an expected utility shaped by gains from progress, losses from possible collision (weighted by risk aversion), and probability distortion:

$$U_i(v, \theta) = (1 - w(p_i)) v(\Delta s_i) - W_c w(p_i)$$

with $v(\Delta s)$ an S-shaped value function and $w(p)$ a nonlinear probability weighting term, as prescribed by Prospect Theory (Haque et al., 2024).

Cognitive Map Navigation and Knowledge Updating

Agent knowledge is encoded as a topological or weighted graph of explored regions, with probabilistic or confidence-ellipse-based locations of the target. Route choice is made via greedy or Dijkstra-style search, with costs updated as new rooms or corridors are discovered, and edges penalized or favored by perceived crowding or density (Andresen et al., 2016, Chraibi et al., 2017).

4. Empirical Validation and Quantitative Performance

Cognitively inspired pedestrian models have undergone rigorous empirical validation against naturalistic datasets:

• Trajectory Reproduction: Hierarchical BT+SFM models achieve sub-2 m STED and FD trajectory error—marked improvement over pure interpolation or rule-based models. Decision accuracy at critical choice points (e.g. crosswalk selection, gap acceptance) reaches 98–100%, validating the high-fidelity replication of real trajectories (Larter et al., 2022).
• Parameter Fitting and Individual Differences: RL-based bounded rationality models adapt parameters such as sensory noise ($\sigma_v$), cognitive weights, and reward tradeoffs to individual experimental subjects, fitting empirical onset-time distributions and reproducing key empirical effects (e.g., speed-dependent gap acceptance, looming aversion, TTA-based decisions) (Wang et al., 2024, Wang et al., 2024, Wang et al., 5 Jan 2026).
• Macroscopic and Crowd Phenomena: Prospect-Theory-based models calibrated on real-world bottleneck, lane formation, and shockwave datasets match observed capacities, lane organization, and density-wave statistics, with RMSE < 10 cm and capacity/jam density measures matching field data (Haque et al., 2024).
• Scenario-based Testing for AVs: Human-like models with inter- and intra-individual variability produce realistic, rare safety-critical interactions (e.g., low-PET, emergency jaywalking). AV controllers optimized using these scenarios attain improved trade-offs between safety (elimination of PET < 1.5 s), comfort (reduction in sudden braking), and efficiency (decreased time lost), compared to overly conservative outcomes when optimizing with rule-based adversarial models (Wang et al., 5 Jan 2026).

5. Range of Cognitive Features and Limitations

A diverse array of cognitive features is already represented:

• Perceptual Limits: Limited field of view, attention and sequential scanning, working memory buffers for vehicles outside current gaze (Kotseruba et al., 2023).
• Individual Preferences: Aggressiveness, law obedience, risk attitude (via loss aversion), type/trait variations (Rasouli et al., 2022, Haque et al., 2024).
• Motor Constraints: Biomechanical, energy-optimal models for stepwise speed selection, effort penalties for acceleration changes (Wang et al., 2024).
• Attention and Distraction: Explicit simulation of distraction (e.g., by reducing decision frequency), yielding observed chaoticities in trajectories (Echeverría-Huarte et al., 2022).
• Dynamic Learning and Adaptation: Event-driven updating of knowledge, opportunistic exploration in ambiguous or unknown environments (Andresen et al., 2016, Chraibi et al., 2017).

Limitations include the current lack of fully automated, data-driven learning for BT structure, force constants, and scenario-dependent calibration; point-mass representations (ignoring body pose or group structure); absence of detailed modeling for implicit negotiation cues (eye contact, subtle deceleration), and simplified treatments of perception and memory noise (Larter et al., 2022, Kotseruba et al., 2023).

6. Comparative Analysis with Classical and Machine-Learned Approaches

Unlike purely force-based or classical discrete-choice models, cognitively inspired frameworks:

• Generate empirically measured phenomena (gap acceptance, lane formation, anticipation, adjustment under visibility or distraction constraints) with clear mapping to cognitive processes.
• Offer interpretability and reusability (e.g., via explicit BTs or parameterized cost/reward functions) lacking in end-to-end deep models or ad hoc rule sets.
• Demonstrate robustness and generalization to untrained scenarios and unseen environments (e.g., simulating rare crossing decisions, route planning in novel buildings) (Larter et al., 2022, Vasishta et al., 2018).
• Provide an interface for scenario-based safety validation that is both realistic and controllable for AV applications (Wang et al., 5 Jan 2026).

7. Prospects and Future Directions

Promising future directions include:

• Automated BT and parameter optimization from trajectory datasets, potentially via hierarchical Bayesian or inverse RL techniques.
• Incorporation of richer perception pipelines, modeling multi-modal uncertainty, occlusion, group interaction, and attentional state.
• Expansion to group, cultural, or demographic diversity in cognitive and decision parameters.
• Application of cognitive risk integration and uncertainty-augmented dynamics to multi-agent simulation in highly interactive AV–pedestrian domains (Dang et al., 28 Jul 2025).
• Scenario-based SOTIF (ISO 21448) testing for AVs, leveraging the fine control enabled by interpretable, cognitively motivated pedestrian agents (Larter et al., 2022, Wang et al., 5 Jan 2026).

This field synthesizes approaches from cognitive science, behavioral psychology, reinforcement learning, and dynamical systems, yielding simulation frameworks that are both descriptive of empirical phenomena and prescriptive for safety-critical engineering applications. Models with explicit cognitive layers now dominate research for high-fidelity, controllable, and interpretable simulation of pedestrian behavior in realistic traffic and crowd contexts.