Plan to Learn, Guard for Ignorance: Contingency MPC that Treats Epistemic and Aleatoric Uncertainty Differently Over Time

A self-driving car that meets an unfamiliar object has two different kinds of doubt about it. The first is noise: blurry pixels, jittery boxes, a pedestrian whose next step is random. This doubt never goes away and grows the further ahead you predict. The second is ignorance: the perception network has never seen this kind of object, or does not know how it moves. This doubt shrinks as the car keeps looking.

Current controllers that model both kinds of doubt treat them the same way through time, and they also assume the object stands still. The result is predictable. The safety margin balloons and the car freezes. Uncertainty-Aware Distributionally Robust Model Predictive Control, the closest existing work, reports a 95% success rate on an out-of-distribution (OOD) motorcycle in CARLA, with the remaining 5% lost to exactly this kind of freezing. Its obstacles are static over the whole prediction horizon.

We propose a controller that plans and actuates in one optimization and handles the two kinds of doubt on different schedules. Noise is spread forward along each obstacle's predicted motion. Ignorance is expected to shrink along the car's own planned trajectory, because the car can predict how much evidence it will gather. To stay safe while being optimistic about learning, the controller commits only to a first control action that also admits a safe fallback under the full, unshrunk ignorance. In two sentences: the car plans as if it will learn more about a strange moving object, and it only takes actions from which it could still stop safely if it learns nothing. This should remove freezing without paying for it in collisions.

Core Claim

Model predictive control (MPC) re-solves a finite-horizon optimal control problem at every time step and executes only the first action. The file calls a controller "unified" when that single optimization chooses steering and braking directly, subject to both obstacle and tire constraints, rather than tracking a path drawn by a separate planner.

Our claim is that, in unified MPC around moving obstacles, freezing and over-conservatism come from one specific modeling error: epistemic uncertainty is propagated as if it were aleatoric. Aleatoric uncertainty (irreducible data noise) and epistemic uncertainty (model ignorance, reducible by more observations) should follow different laws over the horizon:

  • Aleatoric uncertainty grows with the obstacle's predicted motion.
  • Epistemic uncertainty contracts according to the evidence the ego vehicle's planned trajectory will collect.

A two-branch contingency MPC makes the optimistic contraction safe. The two branches share the first control input. The "progress" branch uses the contracted ambiguity sets. The "fallback" branch uses the full current ambiguity and must end in a stopped or otherwise safe state.

Write $u_0$ for the shared first input and $u^p$, $u^f$ for the progress and fallback input sequences. Write $\mathcal{A}_k$ for the ambiguity set at horizon step $k$ built from current evidence, and $\mathcal{A}_k^{+}(u^p)$ for the set predicted after the evidence the progress branch is expected to gather. The problem solved at each step is

$$ \min_{u^p,\,u^f} J(u^p)\ \text{s.t.}\ u^p_0=u^f_0,\ \ \mathrm{CVaR}_{\mathcal{A}_k^{+}(u^p)}[g_k(x^p_k)]\le 0,\ \ \mathrm{CVaR}_{\mathcal{A}_k}[g_k(x^f_k)]\le 0,\ \ x^f_N\in\mathcal{S},\ \ \Big(\tfrac{F_x}{\mu F_z}\Big)^2+\Big(\tfrac{F_y}{\mu F_z}\Big)^2\le 1 . $$

The symbols are:

  • $J$: the tracking and progress cost.
  • $g_k$: the signed collision constraint against the obstacle's predicted position at step $k$.
  • $\mathrm{CVaR}$: conditional value-at-risk, the expected violation in the worst tail of the distribution.
  • $x^p_k$, $x^f_k$: the states under each branch.
  • $\mathcal{S}$: a set of stopped states that are safe with respect to the full reachable set of the obstacle.
  • The last inequality: a friction-ellipse limit on longitudinal and lateral tire forces $F_x, F_y$, with normal load $F_z$ and road friction $\mu$. It applies to both branches, so the fallback is physically executable and not just geometrically clear.

Two properties follow. If the fallback branch was feasible at the last step, its shifted tail remains a candidate now, which gives recursive feasibility of at least the fallback. The car therefore never faces an empty problem, which is the freezing mechanism in the DRO-EDL-MPC paper. And the car keeps moving forward whenever some first action is compatible with both branches.

Research Program

Evidence model. We reuse the evidential perception of DRO-EDL-MPC. That work predicts a Normal Inverse-Gamma distribution per 3D box, with parameters $(\gamma,\nu,\alpha,\beta)$, using MEDL-U trained on KITTI. It scales a distributionally robust CVaR ambiguity set by these parameters, and it avoids mixed-integer programs through a precomputed lookup table of continuous upper bounds. We extend the box state with velocity and heading rate, so the ambiguity describes the obstacle's motion and not only its pose.

The Normal Inverse-Gamma family is conjugate to Gaussian observations. Its evidence counts $\nu$ and $\alpha$ therefore accumulate naturally across frames, and the epistemic variance $\beta/(\nu(\alpha-1))$ shrinks as they do. Consecutive frames are correlated, so naive accumulation over-counts evidence. We learn a small regressor, from logged CARLA and nuScenes sequences, that predicts the realized per-frame evidence gain as a function of range, relative viewing angle, occlusion fraction and weather. $\mathcal{A}_k^{+}$ is computed from this regressor applied along the progress branch's predicted relative poses. The prediction depends on the ego plan, so we solve the problem by sequential convexification: we fix the previous iterate's poses, compute $\mathcal{A}^{+}$, and re-solve.

Aleatoric propagation. Aleatoric motion uncertainty is propagated with a heteroscedastic forecaster, a network that predicts both a mean trajectory and its own noise level, in the style of Long-Term On-Board Prediction of People in Traffic Scenes Under Uncertainty. That paper paired heteroscedastic heads with MC Dropout, a sampling-based way to estimate model disagreement, but ran open-loop and only to 1.0 s. We keep only the single-pass heteroscedastic head, because epistemic uncertainty already comes from the evidential parameters. We recalibrate it with context-conditioned temperature scaling from CCTR. CCTR showed that raw forecasters' variances are nearly uncorrelated with their errors, and it fed calibrated forecasts only to decoupled geometric planners such as RRT*.

Short-horizon worst case. For the first 0.5 s we intersect the prediction with a full physical reachable set, following Safe Planning and Control Under Uncertainty for Self-Driving. That work combined reachability and goal-directed prediction with a robust tube MPC. Its GP-learned disturbance bound lumps all uncertainty together, and its kinematic model cannot see tire saturation.

Vehicle model. The vehicle is a dynamic bicycle model with the friction ellipse above. This is the physical-trackability requirement stressed by the integrated motion planning and tracking survey of Shuyou Yu et al..

Formal framing. The set-valued, prior-free treatment of epistemic uncertainty is borrowed conceptually from Probabilities Are Not Enough. That paper showed that controllers modeling only stochastic noise can drop to a 6% safety rate under parameter mismatch. It synthesizes offline over discrete abstractions of stationary workspaces. We use the same separation of uncertainty types online, in continuous MPC, around moving agents.

Compute target. The goal is a 20 ms loop, matching the 50 Hz Khaitan et al. report, on a laptop CPU. The contingency structure roughly doubles the number of decision variables. The lookup-table CVaR bounds keep each branch a smooth program.

Baselines

Naive reference. A decoupled Frenet planner with a fixed 3 m margin, tracked by a nominal MPC. The DRO-EDL-MPC paper's single-estimate baseline, a closely related non-robust reference, reached a 100% collision rate in its OOD case. We expect a similar failure and include this baseline to anchor the scale.

Strongest existing method. DRO-EDL-MPC. It is the only method in the file that combines hybrid evidential uncertainty with unified MPC, and it reports 95% success, 2% collisions and 5% freezing on the static OOD motorcycle, with a 5× speedup over Wasserstein distributionally robust optimization. We run it in two forms:

  • As published: obstacles static over the horizon.
  • Obvious dynamic extension: constant-velocity obstacle propagation, with the evidential ambiguity carried unchanged or inflated with horizon length.

Both forms receive:

  • the same MEDL-U detector;
  • the same forecaster;
  • the same dynamic bicycle model and friction constraint;
  • the same solver;
  • the same weight-tuning budget, 50 Bayesian-optimization trials on held-out CARLA seeds.

Beating it means lower freeze and timeout rates at equal or lower collision rate, on the moving-obstacle suite, with overlapping or better runtime.

Lumped-uncertainty controller. Khaitan et al.'s reachability-plus-prediction tube MPC, upgraded with the same dynamic tire model so that only the uncertainty treatment differs. Its prediction variance is calibrated with CCTR. This is the strongest method that handles motion but not the epistemic–aleatoric split, and it reports collision-free maneuvers at 20 ms per loop.

Distributionally robust reference. Standard Wasserstein distributionally robust CVaR-MPC, which reached only 29% success in the confident case of the DRO-EDL-MPC paper.

Ablations isolating the key ingredient. The first ablation keeps the two-branch structure but sets $\mathcal{A}^{+}=\mathcal{A}$, so there is no anticipated learning. The second keeps anticipated contraction but removes the fallback branch. The first tells us whether contraction does the work. The second tells us whether the fallback is what keeps it safe.

Why not just…?

Why not just extend DRO-EDL-MPC with constant-velocity obstacle motion? This is the obvious move and the main threat, so it is a headline baseline rather than a straw man. Our prediction is that carrying current epistemic ambiguity along a moving trajectory widens the swept region over the horizon. That should make the static-case freezing more frequent, not less, as sketched in the file's gap analysis. If this extension matches our freeze rate on the moving-obstacle suite, the contribution reduces to the fallback branch alone, and ablation one will show it.

Why not just clip the ambiguity radius at a maximum? Clipping would stop sets from exploding and would likely remove most freezing. It does so by ignoring the very ignorance that made DRO-EDL-MPC safe on OOD objects, where plain CVaR had 44% collisions. We run clipping at three thresholds. It counts as a real alternative if any threshold matches our freeze rate without raising collisions on the OOD subset.

Why not just brake when the MPC becomes infeasible? A braking fallback converts freezing into stopping. It does not restore progress, and in dense traffic a late stop from a planner that never budgeted for it may itself violate friction limits or cause rear conflicts. Our fallback branch is planned in advance and friction-feasible. We report freezes, stops and timeouts separately so the difference shows.

Why not just use a well-calibrated total variance, as CCTR does? Calibration fixes the size of the uncertainty but not how it evolves: a calibrated total variance still cannot tell that it will shrink after two more looks. The lumped-uncertainty baseline tests exactly this. If it matches us, then disentangling adds nothing in the closed loop.

Decisive Experiments

The main suite is CARLA, following Dosovitskiy et al.. We rebuild the DRO-EDL-MPC confident-car and OOD-motorcycle scenarios and add four moving variants:

  • a motorcycle cutting in;
  • a cyclist crossing at sampled angles and speeds;
  • an occluded pedestrian stepping out;
  • a slow OOD vehicle to overtake on a wet road with reduced friction.

Each variant is run for 100 seeds. The metrics are:

  • success rate, collision rate, freeze rate and timeout rate, reported separately;
  • time to goal;
  • minimum separation;
  • peak friction utilization;
  • solve time per step.

We also record the fraction of steps where only the fallback branch was feasible.

Offline, the forecaster's calibration is checked on Argoverse with ECE, the expected calibration error. CCTR reports an 11–22% ECE reduction there. Evidence-regressor transfer is checked on nuScenes and nuScenes-C, the corrupted variant, to cover weather shift.

The first decisive check tests the premise directly: does predicted epistemic contraction match realized contraction in closed loop? We log the predicted $\mathcal{A}^{+}$ against the ambiguity actually observed at the future step. If realized evidence gain is systematically below prediction, the optimism is unfounded. The method then collapses to the fallback branch, and we will report that.

The second decisive check is the freeze-rate comparison against the constant-velocity extension of DRO-EDL-MPC at matched collision rate.

The third is whether the published 5% static-case freezing disappears when our controller is run on the original static OOD scenario. That scenario is the one setting where a reported number, 95% success, exists to beat directly.