Automated cars driving in traffic with human drivers need a safety buffer, a margin that says how hard the human ahead might brake before the automated car must react. Today that buffer is either fixed in advance to cover a worst case, which makes the automated car timid and wastes the traffic-smoothing benefit it was deployed for, or it is fitted to past data in a way that assumes rare extreme maneuvers look like ordinary ones. Human driving does not behave that way. Most of the time acceleration noise is small, but occasionally a driver brakes far harder than a bell curve predicts. Drivers also differ, and hard braking is a different kind of event from accelerating. We propose to estimate, online and separately for each human driver and for the braking and accelerating directions, how heavy the tail of that driver's surprise accelerations is. That estimate then sets the margin inside a cooperative safety controller.
The central argument is about extrapolation. The per-step risk levels a safety controller needs are tiny, on the order of one violation in ten thousand control steps. Over a few seconds of observing a specific driver, no purely empirical quantile can resolve a risk level that small. A Student-$t$ model, a bell-shaped distribution whose "degrees of freedom" parameter $\nu$ controls how slowly its tails decay, can extrapolate from moderate deviations to extreme ones. If $\nu$ is learned rather than fixed, the margin can tighten toward Gaussian levels for a calm driver and widen quickly when a driver starts producing shocks. The new contribution is to connect a learned tail parameter to a control barrier constraint. A control barrier function is a constraint on the controller that keeps the state inside a safe set. We then test whether this connection delivers the target violation rate at lower conservatism than the alternatives.
Core Claim
In a mixed platoon, a cooperative safety layer whose margin on each human driver's braking is set by an online-estimated, direction-specific Student-$t$ tail will do two things. It will hold empirical barrier-violation rates near a prescribed per-step level $\delta$ under realistic heavy-tailed human behavior. And at that same violation rate it will track the desired velocity more closely than three alternatives: offline-calibrated conformal margins, Gaussian margins with inflated variance, and Student-$t$ margins with fixed $\nu$. We expect the advantage to be largest exactly where empirical calibration is weakest: small $\delta$, little per-driver data, and drivers whose behavior shifts mid-episode.
What Motivates, Constrains, and Supplies the Method
Two papers motivate the gap. Enforcing Cooperative Safety for Reinforcement Learning-Based Mixed-Autonomy Platoon Control combines multi-agent reinforcement learning with a differentiable cooperative barrier-function layer, and it bounds human acceleration with conformal prediction calibrated offline in simulation. Its regions are static, so a driver whose tail differs from the calibration population is either over- or under-protected. We keep its controller architecture and change only how the human-uncertainty term is produced. Calibrating car-following models via Bayesian dynamic regression shows on HighD and OpenACC that car-following residuals are autocorrelated and best modeled as an autoregressive process, but it keeps the driving innovation Gaussian and uses the model only offline. We adopt its autoregressive residual structure and replace the Gaussian innovation with an adaptive heavy-tailed one used in closed loop.
Three lines of work constrain the design. Robust worst-case approaches, namely the $H_\infty$ controller of Learning Optimal Robust Control of Connected Vehicles in Mixed Traffic Flow and the robust MPC baseline in Interactive Multi-Modal Motion Planning With Branch Model Predictive Control, show the cost of fixed bounds. In the latter paper, the robust planner stays permanently stuck behind a human vehicle, which is why we refuse to set margins by a fixed maximum disturbance. Student-t Process Regression with Student-t Likelihood shows that learning $\nu$ lets a regression model move between robust and Gaussian behavior. However, its Laplace-approximation inference scales cubically in data size and cannot run at 10–50 Hz, so we need a recursive estimator instead.
Two papers supply the method. Fat-Tailed Variational Inference with Anisotropic Tail Adaptive Flows shows that sharing one tail parameter across dimensions with different tail behavior smears heavy tails onto light ones. We apply the same lesson along two different axes: separate drivers, and the braking versus accelerating sides of each driver's innovation. The Student-$t$ decoders of Student-t Variational Autoencoder for Robust Multivariate Density Estimation supply the bounded log-square likelihood that makes our estimator stable when a single shock arrives.
Research Program
The first component is a per-driver innovation model. For each human vehicle $j$ we predict its acceleration with a calibrated car-following model plus an autoregressive residual of order four to six, following the "sweet spot" reported by Calibrating car-following models via Bayesian dynamic regression. The remaining innovation $\varepsilon_j$ is modeled as a two-piece Student-$t$ with scale $\sigma_j$ and separate tail parameters $\nu_j^{-}$ for braking and $\nu_j^{+}$ for accelerating.
Estimation runs online through the Student-$t$ scale-mixture form. Each new residual $r$ receives a weight $w=(\nu+1)/(\nu+r^2/\sigma^2)$. This weight downweights extreme residuals when updating $\sigma$, while the residuals themselves update a posterior over $\nu$. A forgetting factor discounts old observations, and a population prior fitted on HighD keeps estimates sensible when only a few seconds of a driver's behavior have been seen. Each update costs constant time per step, which fixes the speed problem of the Laplace approximation.
The second component turns the tail estimate into a margin. With $h$ denoting the spacing-based barrier function and $\alpha$ its class-$\kappa$ gain, the cooperative constraint for the automated vehicle following driver $j$ becomes
$$\dot h(x,u;\hat a_j) + \alpha\,h(x) \;\ge\; \gamma\,\sigma_j\, q_{\bar\nu_j^{-}}(\delta),$$
where $\hat a_j$ is the predicted human acceleration and $\gamma$ is the sensitivity of $\dot h$ to that acceleration. The term $q_{\nu}(\delta)$ is the upper-$\delta$ quantile of a unit Student-$t$ with $\nu$ degrees of freedom. The value $\bar\nu_j^{-}$ is a lower posterior credible bound on the braking-side tail parameter, chosen so that uncertainty about the tail makes the margin larger, never smaller.
The resulting margin has a predictable shape. For large $\nu$ it collapses to the Gaussian value $\sigma z_\delta$. For small $\nu$ it grows like $\delta^{-1/\nu}$, which is the polynomial tail decay of the Student-$t$. To guard against misspecification of the tail shape, we add a light online conformal correction. It rescales $\sigma_j$ using the observed exceedance rate at a moderate level such as $10^{-2}$, where empirical counts are informative, and leaves extrapolation to smaller $\delta$ to the parametric tail.
The third component inserts this margin, unchanged otherwise, into the cooperative barrier-function quadratic program of Enforcing Cooperative Safety for Reinforcement Learning-Based Mixed-Autonomy Platoon Control. There the nominal policy comes from multi-agent PPO, and the safety layer shares the burden of a given margin among the cooperating automated vehicles.
Baselines
The naive reference is the same platoon with a Gaussian innovation model whose variance is fit on HighD and inflated until the closed-loop violation rate at $\delta=10^{-4}$ matches our target in the heavy-tail regime. Running at matched safety makes the comparison about how much velocity tracking each method gives up to buy that safety. A second naive point is the nominal multi-agent controller with no safety layer, which that paper reports colliding under sudden braking. We also include a fixed worst-case bound $d_{\max}$ set at the 99.99th percentile of HighD braking innovations.
The strongest existing method is the cooperative barrier function with offline conformal prediction from Enforcing Cooperative Safety for Reinforcement Learning-Based Mixed-Autonomy Platoon Control. It is reimplemented with its conformal set calibrated on the same HighD-derived simulation data our prior uses, the same PPO policy weights, and the same quadratic-program solver. That paper reports nearly 100% safe-region expansion in one scenario and 70% in another relative to the controller without a safety layer. For average absolute velocity error, its efficiency metric, there is no reported number on Lacuna, so we will measure it ourselves under identical disturbances. We also run an adaptive variant that recomputes conformal quantiles over a sliding per-driver window. This is the fairest competitor because it adapts online without any distributional assumption.
The innovation-model baseline is the dynamic IDM from Calibrating car-following models via Bayesian dynamic regression, which reports 5-second acceleration RMSE of 1.66 ± 0.38 versus 2.03 ± 0.48 for memory-augmented IDM. We compare against it on held-out log-likelihood and on extreme-quantile exceedance rates, not on RMSE alone.
The ablations isolate each ingredient. One fixes $\nu=3$. One learns a single $\nu$ shared across drivers. One learns a symmetric per-driver $\nu$. One removes the credible-bound conservatism. One removes the conformal scale correction. A further baseline fits a generalized Pareto tail above a threshold, the standard extreme-value approach. All methods share the same tuning budget: a grid over $\alpha$ and over each method's one or two margin hyperparameters, selected on validation episodes.
Beating the strongest baseline means reaching an empirical per-step violation rate within a factor of two of $\delta$ in all three disturbance regimes, with lower velocity-tracking error than the conformal baselines at that matched rate, and no increase in quadratic-program infeasibility.
Why not just…?
Why not just use adaptive online conformal prediction? This is the most serious threat, because it adapts without assuming a tail shape. Its weakness is resolution. Estimating a $10^{-4}$ quantile from one driver's residuals needs on the order of tens of thousands of samples, or many minutes at 10 Hz. Before that, the window either falls back to a population quantile or under-covers. Our parametric tail extrapolates from far fewer samples, but it pays if the tail shape is wrong. The sliding-window conformal arm, run at $\delta \in \{10^{-2}, 10^{-3}, 10^{-4}\}$ and per-driver data lengths from 10 to 300 seconds, settles this. If conformal matches us at $10^{-4}$ with short windows, the parametric component is not earning its place and we would retreat to the hybrid as a calibration tool only.
Why not just inflate the Gaussian variance or use a fixed worst-case bound? A Gaussian sized to cover rare shocks is oversized for ordinary driving at every step. A fixed bound produced the permanently stuck robust planner in Interactive Multi-Modal Motion Planning With Branch Model Predictive Control. The matched-safety comparison measures this cost directly as velocity-tracking error and throughput.
Why not just fix $\nu=3$? A fixed $\nu$ gets the tail right only for drivers whose tail happens to match it. Fat-Tailed Variational Inference with Anisotropic Tail Adaptive Flows reports a test log-likelihood of −1.659 for a shared tail versus −1.603 for per-dimension tails on CMS claims. That is offline evidence only, and the driver-heterogeneity regime here could differ, so the fixed-$\nu$ and shared-$\nu$ ablations are decisive. If a well-chosen fixed $\nu$ matches the adaptive model on HighD-derived drivers, the honest conclusion is that per-driver adaptation buys nothing on this data.
Why not just fit an extreme-value tail? Peaks-over-threshold methods target exactly extreme quantiles. They need enough exceedances above the threshold, though, and they give no likelihood for ordinary residuals, which our estimator uses to update the scale. We run it as a baseline with the same forgetting and prior.
Proposed Experiments
The offline stage fits innovation models on HighD and OpenACC car-following pairs and evaluates them on held-out drivers. The measures are held-out log-likelihood, and empirical exceedance rates at $10^{-2}$, $10^{-3}$ and $10^{-4}$ as a function of how many seconds of that driver have been seen. We also test whether the estimated braking-side $\nu^{-}$ differs systematically across driving-style groups extracted from NGSIM, following the style labels of Human-Like Decision Making for Autonomous Driving: A Noncooperative Game Theoretic Approach. The specific check is whether aggressive drivers show heavier braking tails than conservative ones.
The closed-loop stage uses the mixed-platoon scenarios of the cooperative-safety paper and the ring and merge environments of Flow: A Modular Learning Framework for Mixed Autonomy Traffic. To avoid testing the model on its own assumptions, human vehicles are driven by IDM plus an autoregressive residual whose innovations are block-bootstrapped from real HighD residuals, not sampled from a Student-$t$. We run three regimes: calm cruising, residuals concentrated around hard-braking events, and a mid-episode switch from a normal to an aggressive driver. The metrics are per-step barrier-violation rate against $\delta$, collisions, minimum time-gap, quadratic-program infeasibility, average absolute velocity error, and ring-road average speed.
The decisive checks are three. First, in the switching regime, our margin must widen within a few seconds of the first shocks, with violation rate staying near $\delta$, while the offline conformal baseline exceeds it. Second, in the calm regime, our tracking error must approach the unconstrained-Gaussian level and fall clearly below the fixed-$\nu$ and worst-case arms. Third, the sliding-window conformal comparison across data lengths must show the parametric tail winning at small $\delta$ and short windows. If it does not, the proposal's central claim is false.
References
- Enforcing Cooperative Safety for Reinforcement Learning-based Mixed-Autonomy Platoon Control
- Calibrating car-following models via Bayesian dynamic regression
- Learning Optimal Robust Control of Connected Vehicles in Mixed Traffic Flow
- Interactive Multi-Modal Motion Planning With Branch Model Predictive Control
- Student-t Process Regression with Student-t Likelihood
- Fat-Tailed Variational Inference with Anisotropic Tail Adaptive Flows
- Student-t Variational Autoencoder for Robust Multivariate Density Estimation
- Human-Like Decision Making for Autonomous Driving: A Noncooperative Game Theoretic Approach
- Flow: A Modular Learning Framework for Mixed Autonomy Traffic