Is Yielding a Trait? Matched Estimates of Driver Disposition for Responsibility Allocation

Controllers that let robots and autonomous vehicles split collision avoidance unevenly already exist. In one line of work each agent carries a "social preference" number that says how much of the avoidance burden it takes, and agents with different numbers get out of deadlocks faster. These numbers are set by hand. When one side of the encounter is a human driver, the controller has to guess the human's number, and nobody has checked whether humans have a stable number at all. It may be that drivers who yield are simply drivers whose road geometry, speed, and right-of-way left them no choice. This note proposes to answer that question from naturalistic driving data using propensity score matching, a standard causal-inference tool that compares like with like. The method compresses the many situational features of an encounter into a single probability, then pairs up encounters that are similar on that probability.

The core test is simple. Take drivers seen in two separate conflicts within one recording. Label each driver by how much burden they took in the first conflict. Then compare their burden in the second conflict against drivers labeled the other way who faced a matched second situation. If the gap survives matching, yielding carries a persistent disposition that an autonomous vehicle could estimate online from one observed encounter and feed into a responsibility-weighted controller. If the gap vanishes, the hand-set preference parameters in current controllers have no behavioral counterpart. Allocation should then be driven by situation alone, as risk-based schemes already do. Either answer changes how asymmetric controllers should be built.

Core Claim

We claim that the share of avoidance burden a human driver absorbs in an interaction splits into two parts. One part is situational necessity: geometry, kinematics, right-of-way, and the neighbors around the pair. The other is a residual disposition. This residual can be identified from observational trajectories only by matching on the situation, because naive comparisons confuse the two. We further claim that, if the residual exists, a one-encounter estimate of it improves prediction of the burden split in the next encounter beyond what situation-only allocation rules capture.

Operationally, for agents $i$ and $j$ in a conflict, the burden share is

$$s_i = \frac{\Delta_i}{\Delta_i + \Delta_j},$$

where $\Delta_i$ is agent $i$'s deviation (integrated speed or path change) from a free-flow reference trajectory fitted on the same agent's non-interacting segments. The quantity $s_i$ plays the same role as the responsibility weight in barrier-function controllers, so an estimate of it can be plugged in directly.

Relation to Prior Work

The controllers that motivate the question come from two papers by Lyu and colleagues. Responsibility-associated Multi-agent Collision Avoidance with Social Preferences builds a psychological preference angle $\phi_i$ into pairwise control barrier constraints, with weight $\omega_i=\cos^2\phi_i$ and $\phi_i+\phi_j=\pi/2$. A control barrier constraint is an inequality on the controls that keeps the system inside a safe set. The paper reports 46.8% better deadlock resolution and 33% shorter completion than symmetric baselines in a six-agent swap. Its preferences are pre-assigned and static, and it names online estimation for humans as open.

Risk-aware Safe Control for Decentralized Multi-agent Systems via Dynamic Responsibility Allocation instead computes $\omega_i=R_j/(R_i+R_j)$ from local tail risk. It reports 100% safety over 50 ramp-merge trials and 4.7% efficiency gain over a fixed 50/50 split. This second scheme is purely situational. Our study asks whether a disposition term adds anything to it, so it is the natural competitor rather than a foil.

Multi-Hypothesis Interactions in Game-Theoretic Motion Planning weights a politeness objective by an intent probability $p$ that is assumed given. Our estimator is a candidate source for that prior. The earlier barrier-certificate work, Safety Barrier Certificates for Collisions-Free Multirobot Systems and PORCA, allocates burden by kinematic capability. That is exactly the kind of situational covariate we must balance before attributing anything to disposition.

The method comes from the matching literature. Reasoning about Political Bias in Content Moderation is the template for our logic. It used matching on justifiable features to show that a raw 4.20% versus 2.35% moderation gap collapsed to about 2.5% for both groups. We ask the same question of yielding: does an apparent behavioral difference survive once justifiable situational differences are balanced? Propensity Score Matching for Causal Inference with Relational Data supplies relational covariates, meaning aggregates over neighbors up to a hop depth. It also supplies a constraint. Matching with too shallow a neighborhood inflated false discoveries, while deeper-than-necessary neighborhoods were tolerated. We therefore default to deep neighborhoods and report sensitivity to depth. That paper also leaves interference between units open. We do not solve it in general; we sidestep it by making the pair, with its neighbors as covariates, the unit.

Causal Confusion in Imitation Learning constrains the design and in part contradicts our premise. It argues that static data cannot separate causes from effects without interventions. We accept this for learning a driving policy. Our target is narrower: a conditional comparison under an explicit no-hidden-confounder assumption. We state that assumption openly and probe it with negative controls and sensitivity analysis rather than claiming it holds.

Research Program

Data and extraction. We extract two-conflict drivers from the INTERACTION Dataset. It has over 40,000 trajectories and over 13,375 interactive pairs, with tight interactions at time-to-conflict of 1 s or less, across roundabouts and merges. Roundabout entry followed by an in-circle conflict naturally gives repeated encounters. As a second source we use the Waymo Open Motion Dataset, whose 20-second scenes over more than 100,000 scenes should contain many agents with two conflicts. How many qualifying drivers exist is unknown. Counting them is the first feasibility milestone, and it decides which dataset carries the main analysis.

Treatment and covariates. The treatment $D$ is whether the driver's first-conflict share was below the median predicted for that situation, that is, whether the driver behaved egoistically in encounter one. The outcome is $s_i$ in encounter two.

The covariates $X$ describe encounter two only:

  • lane geometry and right-of-way;
  • the approach speeds and time-to-conflict of both agents;
  • vehicle size;
  • the partner's own first-encounter behavior when available;
  • relational aggregates (mean, minimum, maximum, spread) over neighbors at increasing hop depth.

We estimate the propensity with gradient-boosted classifiers and use caliper nearest-neighbor matching. We check covariate balance, trim cases without common support, and pair the matching with an outcome model so the estimate stays valid if either model is wrong.

Output of the analysis. The analysis produces a matched effect of first-encounter disposition on second-encounter share. We report it by scene type and as a sensitivity bound: how strong a hidden confounder would have to be to erase it.

Downstream use. If the effect is non-zero, we build an online disposition estimate from one observed encounter. We then combine it with the risk-based weight of Lyu and colleagues as a correction to $\omega_i$ and test it in closed-loop simulation.

Baselines

Unadjusted difference. The naive reference point is the raw difference in second-encounter share between first-encounter egoists and altruists. It shows how much apparent disposition is really situation. By analogy with the moderation audit, we expect a sizable raw gap. Matching "beats" it only in the sense of being the credible number. The raw gap is a diagnostic, not a competitor.

Situation-only allocation rule. The strongest existing method for the downstream question is the risk-proportional weight of Risk-aware Safe Control. Its reported numbers are 4.7% efficiency over a fixed 50/50 split and 100% safety in 50 ramp-merge trials. It is the strongest competitor because it already adapts allocation online with no dispositional assumption. We run it three ways:

  • as a predictor of $s_i$ on held-out INTERACTION encounters, with its risk terms computed from the same observed states;
  • in closed loop, re-implemented on the same ramp-merge and six-agent swap scenarios;
  • in the same closed loop, with human-proxy agents whose dispositions are sampled from our estimated distribution.

Beating it means a statistically reliable reduction in held-out share-prediction error, and in closed loop, shorter completion or fewer deadlocks at unchanged safety. For share prediction on INTERACTION, no reported result was found on Lacuna, so both methods are measured fresh under identical splits.

Fixed 50/50 split and fixed-preference controller. We also run the fixed 50/50 split, and R-SVO with hand-set preferences using its authors' reported 46.8% and 33% gains as the reference.

Flexible situational regressor. A gradient-boosted model of $s_i$ on all encounter-two covariates tests whether a learned situational model already absorbs whatever disposition adds.

Ablations. Three ablations isolate the key ingredients:

  • removing the disposition feature;
  • using ordinary single-unit propensity matching without relational neighbor covariates, to test whether network confounding matters here as it did on Foursquare;
  • replacing encounter-one behavior with a shuffled label from another driver in the same scene type.

All predictors get the same features, the same train/test split by recording location, and the same hyperparameter search budget.

Why not just…?

Why not just regress encounter-two share on situation plus a disposition feature, skipping matching? This is a real threat. With rich covariates and a flexible model, regression could give a similar estimate. We will run it. The matched analysis contributes three things regression hides:

  • an explicit common-support check, since aggressive drivers may never appear in some situations;
  • balance diagnostics;
  • estimates that stay valid if either the propensity model or the outcome model is wrong.

If matched and regression estimates agree and overlap is good, we will say so. The scientific finding, whether disposition exists, stands either way, and matching becomes a robustness check rather than the contribution.

Why not just use risk-based allocation online and ignore disposition? Lyu and colleagues' scheme needs no data and already beats 50/50. If our matched effect is near zero, that is the correct conclusion, and the study will have shown that hand-set preference parameters are unnecessary for human partners. The closed-loop comparison with disposition-sampled human proxies settles whether a disposition correction pays off when dispositions do exist.

Why not just train a large joint trajectory predictor on Waymo data, which implicitly learns who yields? Such a model may predict shares well. But Causal Confusion in Imitation Learning shows that predictors trained on observational data latch onto post-action cues. For example, a decelerating driver "predicts" yielding because deceleration is the yielding. Nor does such a model expose a single interpretable weight a barrier-function controller can consume. We include it only as a predictive upper reference and test whether its errors concentrate on off-support situations.

Decisive Checks

The central result is the matched disposition effect with a sensitivity bound. Three negative controls must pass before we believe it:

  • encounter-one label must show no "effect" on a pre-encounter-two covariate withheld from matching, such as lateral lane offset 5 s before conflict;
  • the effect must disappear in encounters two where the driver faced a forced stop;
  • the shuffled-driver label must give a null result.

Second, relational depth: the estimate should stabilize as hop depth grows, echoing the relational-matching finding that under-specification, not over-specification, is what hurts. Third, in the closed-loop scenarios from the two Lyu papers, the disposition-corrected weight must match or improve on the risk-only weight in completion time and deadlock rate with zero safety violations. A clean null on the first check is a publishable answer in its own right: human yielding in these datasets is situational. That would argue against importing fixed social preference parameters into mixed human–robot allocation.