Credal Collision Bounds from a Single Interactive Trajectory Predictor

A self-driving car does not need a perfect picture of every possible future. It needs to answer one question well: if I take this action, how likely is it that someone ends up in my path? Today's trajectory predictors give a single number for that probability. The number is the same whether the other driver is genuinely undecided, because two outcomes are about equally likely, or whether the predictor has simply never seen a scene like this one. We propose to return an interval instead, a lower and an upper probability of collision for a given ego plan. The interval's position reflects real ambiguity in how others will behave. Its width reflects how little the model knows. The interval is computed after training, from one frozen multi-agent predictor, with a single forward pass plus a one-dimensional convex solve.

The technical idea is to take a recent method for "credal" classification and move it from individual classes to the safety events a planner cares about. A credal set is a set of probability distributions rather than one distribution. That method, Efficient Credal Prediction through Decalibration, perturbs a classifier's output scores (logits) as far as the training likelihood allows. Our observation is that a trajectory predictor's mode head is a classifier. The event "the other agents take a maneuver that intersects my plan" is a union of joint modes. Shifting all logits in that union by one shared scalar turns the upper and lower bound on the event into a one-dimensional convex problem. That sidesteps the non-convexity that stops decalibration from shifting several logits at once. We also make the likelihood budget local: it is measured only on calibration scenes that resemble the current one. Without that, the width would be a fixed function of the base probability and could not track how much support the model has for this scene.

Core Claim

For a mixture-output multi-agent predictor such as HiVT, bounds on the probability of plan-relevant events can be obtained from a group-shifted, locally budgeted decalibration. Those bounds have three properties. First, the interval covers observed event frequencies more reliably than the precise probability. Second, its width rises under distribution shift and under out-of-distribution ego maneuvers, while its midpoint stays stable. Third, it runs within a real-time budget. We expect these event-level bounds to be tighter than summing per-mode credal boxes. We expect them to be at least as informative about missed collision risk as the Dirichlet-based uncertainty of Evidential Uncertainty Estimation for Multi-Modal Trajectory Prediction and the ensemble upper envelope, and much faster than the ensemble.

The Mechanism

Let $z(x)$ be the joint-mode logits a frozen predictor outputs for scene $x$. For ego plan $a$, let $A_a(x)$ be the set of joint modes whose trajectories come within a collision margin of the plan. Write $S_A$ and $S_{\bar A}$ for the summed exponentiated logits inside and outside that set. Adding a scalar $c$ to every logit in $A$ gives

$$P_c(A_a \mid x) = \frac{e^{c} S_A}{e^{c} S_A + S_{\bar A}},$$

which increases monotonically in $c$. We admit every $c$ whose weighted log-likelihood loss on calibration scenes $i$, relative to the unshifted model, stays within the budget $\log \alpha$:

$$\sum_i w_i(x)\,\big[\log p_c(y_i \mid x_i) - \log p_0(y_i \mid x_i)\big] \ge \log \alpha .$$

Here $y_i$ is the observed joint mode in calibration scene $i$. The shifted event $A_{a_i}(x_i)$ is defined on each calibration scene using its logged ego future, so the same event family applies everywhere. The weights $w_i(x)$ are a kernel over the predictor's scene embedding.

Each log-softmax term is concave in $c$, so the feasible set is an interval $[c_{\min}, c_{\max}]$. Plugging its endpoints into the formula above gives the lower and upper probabilities. When the current scene has few nearby calibration neighbours, the weights are small, the constraint loosens, and the interval widens. This localization is the proposal's key ingredient and gets its own ablation. A secondary extension applies the same budgeted shift to a log-scale multiplier on the positional spread of the collision-mode components, so that uncertainty within a mode also enters the bound. We treat that extension as optional.

We read the output as follows. A midpoint near one half with narrow width means genuine negotiation ambiguity: the other driver may yield or may not, and the model knows this. A wide interval means ignorance. The planner would treat the two differently, hedging in the first case and falling back to a cautious maneuver in the second, as in the contingency scheme of LookOut.

Relation to Prior Work

Three works motivate the gap. Entropy-Based Uncertainty Modeling for Trajectory Prediction in Autonomous Driving splits ensemble uncertainty over Gaussian mixtures into an aleatoric part and an epistemic part. It shows the epistemic part rises under sensor blackout and lane deletion while the aleatoric part stays flat. This is evidence that the separation is worth having, but the method needs ensembles. Evidential Uncertainty Estimation gets a single-pass split on a HiVT backbone by predicting Normal-Inverse-Gamma parameters for coordinates and Dirichlet parameters for modes. It commits to one precise second-order distribution, and its heads are per-agent marginals that are not conditioned on an ego plan. Towards a Fatality-Aware Benchmark shows that for reaction prediction the errors that matter are missed collision risk ("non-defensiveness") versus excess caution ("conservatism"). All its models output precise probabilities.

The method comes from decalibration (EffCre). That work is limited to classification and to shifting one logit at a time. We change both the object being bounded (events rather than classes) and the budget (local rather than global). The set-versus-distribution comparison in Set-based v.s. Distribution-based Representations of Epistemic Uncertainty supplies the argument for why a set, not a Dirichlet, is the right object. We test that argument rather than assume it.

Action conditioning comes from Prediction by Anticipation, which predicts others' reactions to a proposed ego maneuver and generalizes to rare maneuvers such as hard braking. We keep its "anticipate, then predict the reaction" structure but attach bounds to the reaction probability. Hardness results for exact credal inference in graphical models (de Campos et al., Mauá et al.) constrain the design. They are why we never build a credal network over agents. We bound only the events the planner queries, and we accept that the bounds are inner approximations of the full credal set.

Research Program

We first fine-tune HiVT to predict a small set of joint modes over the focal agent and its most interactive neighbour, and we condition it on a candidate ego plan, following the anticipation factorization. We then build the event map from plan to colliding joint modes, the embedding-kernel weights, and the one-dimensional solver. Newton's method on a concave scalar function converges in a handful of steps, and the neighbour lookup uses a precomputed index.

The analysis has two parts. The first checks whether the width behaves epistemically: it should grow as calibration support is removed, and it should shrink as calibration data is added back in-distribution. The second checks whether the midpoint behaves aleatorically: it should track the observed frequency of the event in held-out scenes whose conditions match. We also measure how loose the shared-scalar restriction is. On small mode sets, we compare it against the exact event bound obtained by brute-force search over all feasible logit shifts.

Baselines

The naive reference point is the frozen predictor's precise mode probability, temperature-scaled on the calibration split, with the event probability obtained by summing modes. Any interval method must beat it on non-defensiveness at matched conservatism, or it is not worth the complexity.

The strongest existing single-pass competitor is evidential HiVT (Evidential Uncertainty Estimation). Its Dirichlet over modes gives a Beta distribution for any event's probability, so its upper quantile is a directly comparable bound. That paper reports minADE 0.83 on Argoverse 2 against 1.06 for standard HiVT, with 5.6 ms latency. We retrain it on the same joint-mode, plan-conditioned targets, with the same backbone and the same hyperparameter search budget. The quantile level is chosen so that its average width matches ours on in-distribution validation data.

The strongest principled but slow competitor is a five-member deep ensemble, with the upper and lower envelope of member event probabilities taken as the interval, following Entropy-Based Uncertainty Modeling. The file reports 62 ms for deep ensembles against 54 ms for MC Dropout (Evidential Uncertainty Estimation). Its members get the same total training compute as our backbone multiplied by five, so the ensemble receives more compute than we do.

The key ablations are the following. Summing per-mode EffCre boxes isolates the value of event-level shifting. A global rather than local budget isolates the value of localization. For fatality-aware reference points on the NGSIM ramp-merge task, Towards a Fatality-Aware Benchmark reports MDN at fatality-aware Brier $B_c = 0.2361$ with non-defensiveness $D = 0.1303$, and IRL at $B_c = 0.2053$ with $D = 0.0698$ but the worst plain Brier, 0.1821.

Beating the baselines means the following. At matched in-distribution width, we should have lower non-defensiveness than evidential HiVT. We should come within noise of the ensemble on out-of-distribution detection from width. We should do both at under a quarter of the ensemble's latency. And we should leave minADE unchanged, which holds by construction because the method is post-hoc.

Why not just…?

Why not just use a deep ensemble's envelope? An ensemble envelope is a legitimate credal set and is the real accuracy threat. If it clearly dominates us on coverage and out-of-distribution detection, our contribution shrinks to a speed trade-off. The experiments report both axes, so a reader can see whether we sit on the accuracy–latency frontier or below it. The file's 62 ms figure already exceeds the roughly 20 ms that HiVT itself needs.

Why not just take an upper quantile of the evidential Dirichlet? It is cheap, needs one pass, and already yields event bounds. The case against it is that a learned precise hyper-distribution can be confidently wrong under shift, because nothing forces its concentration down on unfamiliar scenes, as argued in Set-based v.s. Distribution-based Representations. The decisive check is the Blackout and LaneDeletion corruptions together with held-out ego maneuvers. If the Beta quantile widens under these shifts as much as our interval does, we will say so, and recommend the evidential route as simpler.

Why not just apply EffCre to the mode head and sum the class bounds? This is the obvious "already done" combination, and it is included as an ablation. Summing per-class upper bounds over a union of modes can exceed the true event upper probability. The sum also ignores that raising one mode lowers the others. We expect it to be conservative by a margin that grows with the number of colliding modes, and we measure that margin directly.

Why not just conformalize the collision probability? Conformal prediction is a method that turns calibration errors into sets with a guaranteed coverage rate. It gives marginal coverage under exchangeability, meaning the test scene must be interchangeable with the calibration scenes. An out-of-distribution ego maneuver breaks exactly that assumption. A single conformal set also does not separate ambiguity from ignorance. We will still run split-conformal intervals on the event probability as an additional reference and report their coverage under the shifted test sets.

Proposed Experiments and Decisive Checks

The first experiment uses the Argoverse 2 motion forecasting data (Argoverse 2), with the event defined relative to the logged autonomous-vehicle future. We add the Blackout and LaneDeletion corruptions of Entropy-Based Uncertainty Modeling, applied here rather than to nuScenes. We report three things. The first is how often the empirical event frequency falls inside the interval, in bins of the interval's midpoint. The second is the area under the ROC curve (AUROC) when interval width is used to separate clean scenes from corrupted ones. The third is latency on one GPU.

The second experiment uses the NGSIM ramp-merge setting of Towards a Fatality-Aware Benchmark. The planner acts on the upper probability: it takes the defensive prototype when that upper probability exceeds a threshold. We report conservatism, non-defensiveness, and fatality-aware Brier against the published HMM, MDN, and IRL figures.

The decisive check is a held-out-maneuver split. We remove all calibration and training scenes in which the ego vehicle brakes hard or merges aggressively, then test on those scenes. The proposal succeeds if two things happen. Width must rise sharply on these scenes, and more than the evidential quantile rises. The midpoint must stay about as accurate as on in-distribution scenes, so that the extra caution comes from width and not from a biased estimate.

The proposal fails if the local-budget width is no more predictive of shift than the global-budget width. In that case the width is only a function of how confident the base prediction is, not a measure of ignorance.