Trust-Aware Game-Theoretic Lane-Changing (TGLD)

• The paper introduces a trust-aware coalition game framework that integrates real-time trust evaluation with strategic lane-changing decisions for automated vehicles.
• It dynamically updates trust levels from observed human driving behaviors to adjust cooperative responses during on-ramp merging.
• Validation experiments reveal improved efficiency and comfort, with reduced jerk and enhanced average speeds compared to baseline models.

Searching arXiv for the primary TGLD paper and a few closely related lane-changing/game-theoretic works to ground the article. {"query":"arXiv (Pan et al., 14 Jul 2025) Trust-Aware Game-Theoretic Lane-Changing Decision Framework for Automated Vehicles in Heterogeneous Traffic", "max_results": 5} {"query":"trust-aware game-theoretic lane-changing decision arXiv", "max_results": 10} {"query":"(Pan et al., 14 Jun 2025) Trust-MARL on-ramp merging trust-triggered game-theoretic lane-changing arXiv", "max_results": 5} Trust-Aware Game-Theoretic Lane-Changing Decision (TGLD) is a lane-changing decision framework for automated vehicles in heterogeneous traffic that combines online trust estimation of human-driven vehicles with game-theoretic maneuver selection. In the formulation reported for highway on-ramp merging, the subject vehicle is an automated vehicle, automated vehicles are modeled as fully cooperative, human-driven vehicles are modeled as partially cooperative through real-time trust evaluations, and the decision objective includes safety, efficiency, comfort, and social compatibility, including minimizing disruption to surrounding vehicles and improving the predictability of automated-vehicle behavior (Pan et al., 14 Jul 2025).

1. Definition and interaction setting

TGLD is designed for mixed traffic in which automated vehicles and human-driven vehicles coexist and interact strategically during lane-changing, especially at bottlenecks such as highway on-ramp merging. The central difficulty is not only geometric gap acceptance, but also the fact that human drivers are heterogeneous, may be aggressive or defensive, and may change their willingness to cooperate as they observe the automated vehicle. In this setting, trust is treated as dynamic rather than fixed, and the automated vehicle adapts its lane-changing strategy accordingly (Pan et al., 14 Jul 2025).

The reported TGLD architecture is organized around perception and state estimation, driving-style recognition for human-driven vehicles, an online trust evaluation module, a game-theoretic decision module, a trajectory planner, and low-level vehicle control. Inputs include positions, velocities, accelerations, lane topology, and historical motion of surrounding vehicles. Outputs are a high-level lane-changing decision, associated cooperative maneuvers, and updated trust estimates for relevant human-driven vehicles. The formulation is specific to a subject vehicle attempting to merge from an on-ramp to the mainline, but its design logic is more general: interaction is modeled as a strategic decision problem rather than as isolated gap checking (Pan et al., 14 Jul 2025).

A common misconception is that TGLD is merely a politeness heuristic layered on top of a conventional planner. In the reported formulation, trust affects the interaction model itself: it changes how much cooperation is expected from human-driven vehicles, how their contribution is counted in coalition value, and which maneuvers are considered context-appropriate. Safety, however, remains an explicit objective rather than an emergent by-product.

2. Coalition-game formulation

The formal game is defined over the set of vehicles in the cooperation area, $N = \{V_1,\ldots,V_n\}$. For any non-empty subset $Q \subseteq N$, the coalition is partitioned into automated vehicles $Q_{\text{AV}}$ and human-driven vehicles $Q_{\text{HV}}$. The characteristic function is given by

$$v(Q) = \sum_{i \in Q_{\text{AV}}} u_i + \sum_{j \in Q_{\text{HV}}} p_j\,u_j,$$

where $u_i$ is the payoff of vehicle $i$ and $p_j \in [0,1]$ is the trust level of human-driven vehicle $j$ toward automated vehicles (Pan et al., 14 Jul 2025).

This construction makes the asymmetry between automated and human-driven participants explicit. Automated vehicles are fully cooperative inside the coalition value, whereas human-driven vehicles contribute through trust-weighted partial cooperation. Group rationality is expressed as

$\sum_{i=1}^{m} u_i = v(Q),$

and the exposition also states that individual rationality must hold so that no member has incentive to leave the coalition (Pan et al., 14 Jul 2025).

Operationally, multi-vehicle conflicts are resolved through pairwise lane-changing interactions inside the coalition structure. The framework description treats the relevant merge conflict as the strategic core and evaluates candidate maneuvers by computing per-vehicle payoffs, coalition value, and rationality conditions. This suggests that TGLD is best understood as a trust-weighted coalition game with local conflict resolution, rather than as a fully centralized optimizer over the entire traffic scene.

3. Online trust evaluation

In TGLD, trust is represented by a scalar $Q \subseteq N$0 for each relevant human-driven vehicle. Conceptually, it represents that driver’s willingness to rely on and cooperate with the automated vehicle. The reported trust module updates $Q \subseteq N$1 online from observed behavior, with initial $Q \subseteq N$2, so trust is neutral at the beginning of the interaction and then shifts as evidence accumulates (Pan et al., 14 Jul 2025).

The trust update pipeline uses longitudinal motion as the most informative observable during merging. A longitudinal acceleration model,

$Q \subseteq N$3

is combined with an LSTM-based prediction model

$Q \subseteq N$4

Monte Carlo trajectory sampling is then used to generate candidate futures for the human-driven vehicle. Each sampled trajectory is labeled as forward, backward, or uncertain according to position and velocity thresholds $Q \subseteq N$5 and $Q \subseteq N$6; the empirical proportions of these labels are used to update trust (Pan et al., 14 Jul 2025).

This formulation makes trust a behavioral estimate rather than a psychological abstraction. It does not attempt to infer a rich latent mental model of the human driver. Instead, it asks whether the driver’s recent motion implies a stable, interpretable intention and whether that intention is compatible with cooperative merging. This suggests that, within TGLD, trust is closer to an online estimate of interaction reliability than to a static personality trait.

4. Payoffs, driving style, and social compatibility

The payoff structure distinguishes between human-driven vehicles and the subject automated vehicle. For a human-driven vehicle, the payoff combines safety, efficiency, and comfort, modulated by a driving-style factor $Q \subseteq N$7: $Q \subseteq N$8 The reported style settings are $Q \subseteq N$9 for aggressive, $Q_{\text{AV}}$0 for normal, and $Q_{\text{AV}}$1 for defensive driving (Pan et al., 14 Jul 2025).

For the subject vehicle, the payoff adds a trust-social term: $Q_{\text{AV}}$2 The trust-social payoff is further decomposed as

$Q_{\text{AV}}$3

The first component is defined to minimize the impact on surrounding vehicles through the probability of a favorable response from the front vehicle under the subject vehicle’s maneuver, while the second increases self-predictability by rewarding a larger probability gap between the best and second-best actions (Pan et al., 14 Jul 2025).

This payoff design operationalizes “social compatibility.” It is not an informal appeal to human-friendliness. Instead, it is encoded through explicit reward terms that discourage forcing harsh reactions from surrounding drivers and encourage automated-vehicle behaviors that are easy to interpret. The formulation therefore treats predictability as a control objective rather than as a post hoc evaluation criterion.

5. Decision workflow and maneuver synthesis

The reported decision procedure begins with perception and prediction inside a collaboration area around the merge. In the human-in-the-loop setup, the subject vehicle’s perception range is $Q_{\text{AV}}$4; all vehicles except the human-driven one start at $Q_{\text{AV}}$5; and the initial distance between the two mainline automated vehicles is $Q_{\text{AV}}$6 (Pan et al., 14 Jul 2025).

From the current traffic state, the system first recognizes the human driver’s style, then updates trust, then generates candidate high-level maneuvers for the subject vehicle, including merge timing and longitudinal adjustment. For each candidate maneuver, the framework identifies the relevant coalition, computes vehicle payoffs, computes the coalition value $Q_{\text{AV}}$7, and checks rationality and safety feasibility. The final maneuver is selected subject to safety constraints such as minimum time headway, minimum distance, and dynamic limits on acceleration and jerk, after which a trajectory planner produces a continuous motion plan for execution (Pan et al., 14 Jul 2025).

This staged workflow clarifies an important property of TGLD: trust does not directly output control. It changes the strategic evaluation of candidate maneuvers. High trust permits the planner to rely more on cooperative responses from a human-driven vehicle; low trust makes the subject vehicle more conservative because the coalition value of that interaction is reduced and the expected benefit of assertive merging declines.

6. Human-in-the-loop validation

The reported validation uses a human-in-the-loop experiment in a highway on-ramp merging scenario with 36 participants, comprising 18 aggressive and 18 defensive drivers. Safety is evaluated through minimum time headway and minimum distance, efficiency through average speeds, and comfort through maximum jerk. The cases are grouped as defensive-high-trust, defensive-low-trust, aggressive-high-trust, and aggressive-low-trust (Pan et al., 14 Jul 2025).

For safety, the defensive-high-trust case shows the highest and most stable margins, with median minimum time headway of approximately $Q_{\text{AV}}$8 and median minimum distance of approximately $Q_{\text{AV}}$9. The defensive-low-trust case shows greater variability, with minimum time headway down to $Q_{\text{HV}}$0 and minimum distance around $Q_{\text{HV}}$1. The aggressive-high-trust case remains relatively concentrated, whereas the aggressive-low-trust case exhibits lower safety indicators and stronger defensive adaptation by the subject vehicle (Pan et al., 14 Jul 2025).

For efficiency, the defensive-high-trust case reports average human-driven-vehicle speed of approximately $Q_{\text{HV}}$2 and subject-vehicle speed of approximately $Q_{\text{HV}}$3. In the aggressive-high-trust case, human-driven-vehicle speed lies in the $Q_{\text{HV}}$4–$Q_{\text{HV}}$5 range while the subject vehicle reduces speed to approximately $Q_{\text{HV}}$6. In the aggressive-low-trust case, the subject vehicle slows more substantially, indicating that low trust in aggressive drivers pushes the planner toward defensive strategies (Pan et al., 14 Jul 2025).

For comfort, the defensive-high-trust condition yields the lowest jerk, while aggressive-low-trust interactions produce the highest observed jerk, reaching $Q_{\text{HV}}$7. Across all four conditions, the subject vehicle’s maximum jerk remains lower than the human-driven vehicle’s, indicating smoother motion from the automated controller (Pan et al., 14 Jul 2025).

The ablation study compares TGLD with a baseline game-theoretic lane-changing model without trust awareness. Average speed is significantly higher with TGLD, with $Q_{\text{HV}}$8, and maximum jerk is significantly lower, with $Q_{\text{HV}}$9. Minimum time headway does not differ significantly, with $$v(Q) = \sum_{i \in Q_{\text{AV}}} u_i + \sum_{j \in Q_{\text{HV}}} p_j\,u_j,$$0, and zero collisions are reported in both variants. This suggests that, in the reported study, the trust mechanism improved efficiency and comfort while maintaining safety rather than trading safety away for assertiveness (Pan et al., 14 Jul 2025).

7. Relation to adjacent research and debates

TGLD belongs to a broader line of game-theoretic interactive driving models that move from fixed, non-trust interaction assumptions toward adaptive social reasoning. Earlier work on highway and merging scenarios formulated lane changing as Nash or Stackelberg interaction with safety, comfort, and efficiency costs, but without explicit dynamic trust; examples include a human-like noncooperative framework using Nash and Stackelberg games (Hang et al., 2020) and a work-zone lane-changing planner that models a two-player non-cooperative dynamic game over candidate trajectories (Nour et al., 20 Mar 2026). TGLD differs by making trust a state variable that changes how much partial cooperation is expected from human-driven vehicles.

A closely related development is the trust-triggered game-theoretic lane-changing module inside Trust-MARL, where pairwise trust $$v(Q) = \sum_{i \in Q_{\text{AV}}} u_i + \sum_{j \in Q_{\text{HV}}} p_j\,u_j,$$1 is updated as

$$v(Q) = \sum_{i \in Q_{\text{AV}}} u_i + \sum_{j \in Q_{\text{HV}}} p_j\,u_j,$$2

and a trust threshold switches interactions between cooperative and non-cooperative game modes (Pan et al., 14 Jun 2025). That framework is reinforcement-learning-based and emphasizes repeated coordination at bottlenecks, whereas TGLD is centered on coalition-game lane-change choice with explicit social-compatibility payoffs. The two approaches are complementary rather than contradictory.

Another adjacent line interprets robustness of a game solution to cost-function mismatch as a quantitative trust surrogate. In that formulation, the robustness margin

$$v(Q) = \sum_{i \in Q_{\text{AV}}} u_i + \sum_{j \in Q_{\text{HV}}} p_j\,u_j,$$3

measures how much deviation in other agents’ cost functions can be tolerated without changing the equilibrium (Liu et al., 2023). This suggests a different meaning of trust: not willingness to cooperate, but confidence that the modeled strategic response remains valid under uncertainty.

A more controversial contrast appears in eHMI-mediated lane-changing negotiation, where Bayesian inference is also used to track trust, but strategic signaling may include benevolent deception. That line reports efficiency gains in up to $$v(Q) = \sum_{i \in Q_{\text{AV}}} u_i + \sum_{j \in Q_{\text{HV}}} p_j\,u_j,$$4 of scenarios, safety gains in up to $$v(Q) = \sum_{i \in Q_{\text{AV}}} u_i + \sum_{j \in Q_{\text{HV}}} p_j\,u_j,$$5, and trust collapse in up to $$v(Q) = \sum_{i \in Q_{\text{AV}}} u_i + \sum_{j \in Q_{\text{HV}}} p_j\,u_j,$$6 of drivers in the most pronounced cases (Peng et al., 27 Nov 2025). TGLD itself is framed around social compatibility and adaptive cooperation rather than deceptive signaling, but this comparison shows that trust-aware lane-changing is also an ethical design problem. The field is therefore divided not over whether trust matters, but over how trust should be represented, how aggressively it should be exploited, and what constraints should govern strategic interaction with human drivers.