Pith. sign in

REVIEW 3 major objections 5 minor 19 references

A priority-biased real-time planner can match human safety and regulatory compliance without imitating the human path, once scoring stops rewarding log resemblance.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-14 07:55 UTC pith:NDCQX7L7

load-bearing objection Solid systems paper: a carefully scoped tiered-slack NMPC plus a clean fix for log-coupled scoring bias; simulation-only and soft Tier-1 safety are disclosed limits, not hidden ones. the 3 major comments →

arxiv 2607.10975 v1 pith:NDCQX7L7 submitted 2026-07-13 cs.RO math.OC

Real-Time Rulebook-Aware Nonlinear MPC for Autonomous Driving with Priority-Biased Tiered Slacks

Hadi Hajieghrary , Benedikt Walter , Chaitanya Shinde , Paul Schmitt , Miguel Hurtado This is my paper
classification cs.RO math.OC
keywords model predictive controlautonomous drivingmotion planningrulebooksreal-time optimizationconstraint handlinglog-independent evaluationtiered slacks
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Autonomous driving planners must trade off safety, traffic rules, comfort, and efficiency every control cycle and leave a readable record of which rule gave way. This paper presents W-SQP, a nonlinear model predictive controller that folds nine rule families into four shared-slack tiers with strongly separated quadratic penalties, keeps physical actuation bounds hard, replans from the executed state at 10 Hz, and logs threshold-normalized residuals every cycle. Under a 90 ms solver-time limit the solver returns a best iterate that is projected through the vehicle model before execution, with median and maximum observed solve times of 28 and 104 ms. The second contribution is an evaluation protocol that splits closed-loop metrics into log-independent safety and regulatory rules versus rules that measure resemblance to the recorded human trajectory. On 150 paired Waymo scenarios, W-SQP shows no systematic group-level deficit versus expert replay on the sixteen log-independent rules (mean difference about -0.04 percentage points), while the two log-coupled rules account for a roughly 10-point gap that ordinary aggregates would misread as worse driving. The paper therefore offers both an auditable, priority-biased, anytime-capable controller prototype and a scoring discipline that separates driving quality from imitation.

Core claim

W-SQP, a four-tier shared-slack nonlinear MPC that biases residual violations toward lower-priority rules while keeping actuation bounds hard, reaches group-level parity with expert replay on log-independent safety and regulatory rules across 150 closed-loop scenarios, once those rules are scored separately from two log-coupled imitation metrics that by construction award 100 percent to the recorded human.

What carries the argument

W-SQP (weighted soft-constrained quadratic-penalty NMPC): nine rule families map onto four shared-slack tiers whose quadratic penalties are separated by orders of magnitude (safety ≻ regulatory ≻ comfort ≻ efficiency), solved jointly online so residual violations are biased—not lexicographically forced—toward lower tiers, with hard physical bounds and a per-cycle residual audit log.

Load-bearing premise

The parity and timing claims rest on a simulator with perfect state, simplified vehicle dynamics, and reactive agents that never face real sensing, estimation, or actuation error; if those idealizations fail on the road, both claims can break without any change to the optimizer itself.

What would settle it

Re-run the same 150 paired scenarios (or a matched vehicle-in-the-loop set) with sensing and actuation latency and a non-IDM reactive agent model; if the log-independent safety and regulatory group mean difference versus expert replay then falls systematically outside the reported bootstrap interval of about [-1.1, +0.5] percentage points, the central parity claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper presents W-SQP, a rulebook-aware nonlinear MPC that maps nine driving-rule families onto four shared-slack tiers with strongly separated quadratic penalties, solved online with CasADi/IPOPT. Actuation bounds remain hard; soft constraints are priority-biased rather than lexicographic. The controller replans at 10 Hz from the executed state, logs per-rule residuals, and under a 90 ms CPU-time limit returns an anytime iterate projected through the dynamics (median/max wall-clock 28/104 ms). On 150 paired WOMD/Waymax scenarios it is compared with reactive and proposal-and-select baselines. A supporting contribution is a log-independent evaluation protocol that separates safety/regulatory compliance from log-coupled imitation metrics; under that protocol W-SQP shows a mean difference of −0.04 pp versus expert replay on the sixteen log-independent rules (bootstrap CI [−1.07, +0.52] pp) and −10.1 pp on the two log-coupled rules, with localized safety regressions in high-divergence scenes. The authors carefully scope the result as an auditable, anytime-capable prototype rather than a hard-real-time or formally safe controller.

Significance. If the scoped claims hold, the work makes two useful contributions for autonomous-driving planning and benchmarking. First, it integrates established weighted-slack NMPC primitives into a single real-time, auditable program with explicit priority bias and per-cycle residual logging—practically relevant for post-hoc review of objective conflicts. Second, the log-independent taxonomy and imitation-premium decomposition (Prop. 1–2, Corollary 1) cleanly separate driving quality from resemblance to a recorded trajectory; the empirical decomposition on n=150 paired rollouts, with tier ablations, baseline generality, and density/threshold/selection controls, is a concrete, falsifiable demonstration of a Goodhart-type confound that many closed-loop catalogues risk. Strengths include careful claim scoping, statistical reporting (Wilcoxon + BH, bootstrap CIs), and transparent disclosure of soft Tier-1 safety and ℓ2 (not exact ℓ1) design choices. The result is incremental rather than foundational, but the evaluation protocol is transferable and the controller characterization is reproducible within the stated simulation setting.

major comments (3)
  1. [Sec. 3.4–3.5, Table 1, Table 4] Sec. 3.4–3.5 and Table 1: twelve of the evaluator’s 25 rules (including stop-sign, crosswalk yield, yield-to-priority, lane intrusion) have no dedicated NLP surrogate. The paper notes incidental satisfaction or rare applicability, but the group-level log-independent parity claim (Sec. 7.2, Table 4) then rests partly on rules the controller never optimizes. A clearer breakdown of which R_inv rules are actively constrained versus only evaluated would strengthen the causal link between the tiered program and the reported compliance.
  2. [Sec. 3.2–3.3, Sec. 7.1, Table 4, Fig. 9] Sec. 3.2–3.3 and Sec. 7.1: the priority bias is demonstrated via flat-weight, priority-ratio, and uniform-stiff ablations, which is good. However, residual Tier-1 soft violations remain possible by design (sub-100% safety compliance in Table 4; collision/road-edge relaxation frequencies in Fig. 9). The abstract and conclusion correctly avoid a formal-safety claim, but the phrase “no systematic group-level deficit” can still be misread as safety parity. Explicitly bounding the worst-case Tier-1 residual under the chosen ρ1 (or reporting the distribution of max collision residual) would make the soft-safety design consequence quantitative rather than only qualitative.
  3. [Sec. 6, Sec. 8.4] Sec. 6 and Sec. 8.4: closed-loop validity rests on ground-truth state, IDM agents, and a kinematic bicycle with no sensing, estimation, or actuation latency. The paper discloses this, yet the anytime timing numbers (28/104 ms) and the parity claim are both platform- and model-dependent. At minimum, a short sensitivity discussion—or a statement that the log-independent protocol itself is the transferable contribution even if absolute compliance shifts under more realistic agents—would better separate the methodological claim from the simulation-specific numbers.
minor comments (5)
  1. [Abstract, Sec. 3] The name W-SQP is repeatedly clarified as not denoting sequential quadratic programming; a single early footnote is enough—later repetitions (abstract, Sec. 3) can be shortened.
  2. [Fig. 7, Fig. 12] Fig. 7 and Fig. 12: the β-sweep is a useful continuum, but the caption should state more clearly that blended poses are not bicycle-feasible (as noted in the text), so the curve is a divergence parameterization rather than a family of realizable plans.
  3. [Table 2] Table 2: units of the tier penalties ρj are not commensurate across constraint families (Sec. 3.6 already notes this); a brief reminder in the table caption would help readers interpret the 10^8…1 ladder.
  4. [Sec. 5.2, Table 4] Sec. 5.2: compliance normalizes over all T steps rather than applicable steps only; the relative paired analysis is sound, but a one-sentence reminder near Table 4 would reduce misreading of absolute expert-replay rates (e.g., lateral clearance ~55%).
  5. [Sec. 3.2] Minor typography: “trade-offis” / “trade-off” spacing inconsistencies appear in Sec. 3.2 and elsewhere; a pass for compound-word spacing would clean the text.

Circularity Check

0 steps flagged

No significant circularity: the log-independent evaluation protocol is designed to break imitation-by-construction scoring, and the controller results are empirical closed-loop measurements against external baselines.

full rationale

W-SQP is a design-and-evaluate paper, not a first-principles derivation. The NLP (Eqs. 1–5), tier map, and hand-chosen ρj are disclosed design choices; the central claims are empirical: closed-loop compliance on 150 paired WOMD/Waymax scenarios, solve-time CDFs, and ablations (flat weights, priority-ratio sweep, uniform-stiff, baselines, density/threshold/selection controls). Expert replay is an external baseline that maximizes log-coupled rules by construction (Prop. 1), which the paper uses to motivate separating Rlog from Rinv rather than to fit or force W-SQP’s scores. Compliance is measured by an evaluator geometry independent of the planner surrogates (Sec. 5.2, 3.5). There is no fitted parameter renamed as a prediction, no uniqueness theorem imported from the authors, and no self-citation that carries the load-bearing argument. The paper is self-contained against the stated external benchmarks within its scoped simulation setting.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 2 invented entities

The controller claim rests on standard NMPC/soft-constraint machinery plus hand-chosen tier weights and surrogate geometries; the evaluation claim rests on simulator and metric-construction assumptions (GT state, IDM agents, referenced-state taxonomy). No new physical entities. Free parameters are design knobs (ρ tiers, buffers, horizon) fixed a priori across scenarios rather than fitted to the parity outcome. Domain assumptions about closed-loop simulation fidelity are the main external load-bearing premises.

free parameters (7)
  • Tier slack weights (ρ1,ρ2,ρ3,ρ4)
    Hand-set to (1e8, 1e5, 1e2, 1); separation by orders of magnitude is the priority-bias mechanism. Not fitted to evaluation outcomes, but the quantitative bias depends on these choices.
  • Horizon N and Δt
    N=30, Δt=0.1 s (3 s horizon) fixed; determines NLP size and planning look-ahead.
  • Collision buffer d_safe, road margin, route deviation d_route
    Surrogate geometry thresholds (3.0 m, 0.5 m, 2.0 m) chosen by design; affect soft Tier-1/2 activation.
  • Comfort/jerk/lat-acc and headway thresholds
    Comfort decel 3.0 m/s², lat acc 3.0 m/s², jerk 5.0 m/s³, Thw=1.5 s—design limits for Tier-3/4 surrogates.
  • Base tracking weights Q,R,S and α_T
    Low-weight log reference and input penalties; deliberately small so log alignment sits below safety/regulatory tiers.
  • Solver-time limit 90 ms and IPOPT tolerances
    Anytime behavior and reported 28/104 ms stats depend on max_cpu_time and tol settings.
  • Evaluator strictness κ=1.25 and log-coupled thresholds
    Global strictness and path/route thresholds define compliance scores; sensitivity sweep is reported but nominal values are design choices.
axioms (6)
  • domain assumption Kinematic-bicycle dynamics with RK4 discretization adequately represent ego motion for closed-loop scoring.
    Sec. 3.1 and Sec. 6; no dynamic bicycle, tire, or tracking-error model.
  • domain assumption Strongly separated quadratic slack penalties bias residuals toward lower tiers without providing lexicographic priority.
    Sec. 3.2–3.3; authors state this explicitly and choose ℓ2 for numerical robustness over exact ℓ1 penalty theory.
  • domain assumption Waymax IDM reactive agents plus WOMD logs are a valid closed-loop testbed for relative planner comparison.
    Sec. 6; agent-realism literature the paper cites shows rankings can shift under learned agents.
  • ad hoc to paper A metric is log-coupled iff its predicate references the recorded log; log-independent metrics depend only on executed driving.
    Definitions 1–2 in Sec. 5.3; the fair-evaluation protocol is built on this taxonomy.
  • ad hoc to paper Shared one-slack-per-tier relaxation is an acceptable tractability trade-off (tracks most binding constraint in tier).
    Sec. 3.3, 3.6; enables 4N=120 slacks and 10 Hz, but couples within-tier constraints.
  • standard math Standard interior-point NLP (IPOPT/MUMPS) with warm starts solves the nonconvex program sufficiently often for the study.
    Sec. 4; 0.58% open-loop fallback; anytime iterate projected through dynamics.
invented entities (2)
  • W-SQP (weighted soft-constrained quadratic-penalty / tiered-slack NMPC architecture) no independent evidence
    purpose: Compile a multi-family driving rulebook into one online four-tier shared-slack NLP with residual audit and anytime execution.
    Named system integrating known soft-MPC and rulebook ideas; not a new physical entity. Independent evidence is the closed-loop and timing experiments in the paper; no external deployment evidence yet.
  • Log-independent vs log-coupled metric taxonomy and imitation-premium decomposition independent evidence
    purpose: Separate driving-quality compliance from resemblance to a recorded human trajectory in closed-loop scoring.
    Evaluation construct (Defs. 1–3, Prop. 1–2). Falsifiable on other catalogs/planners; paper already applies it to PP+IDM and PDM-style baselines.

pith-pipeline@v1.1.0-grok45 · 33770 in / 4474 out tokens · 39040 ms · 2026-07-14T07:55:11.036888+00:00 · methodology

0 comments
read the original abstract

Autonomous-vehicle motion planners must resolve conflicts among safety, regulation, comfort, and efficiency in real time while exposing those decisions for audit. We present W-SQP, a weighted tiered-slack nonlinear model predictive controller (NMPC) that compiles nine driving-rule families into a four-tier shared-slack nonlinear program solved online with CasADi and IPOPT; the name denotes the weighted quadratic slack penalty, not a sequential-quadratic-programming solver. Strongly separated tier penalties bias residual violations toward lower-priority rules while leaving actuation bounds hard. The controller replans from its executed state at $10$\,Hz and records per-rule residuals on every cycle. A $90$\,ms solver-time limit returns an anytime iterate that is projected through the vehicle dynamics before execution; median and maximum observed wall-clock solve times were $28$ and $104$\,ms. We evaluate W-SQP in closed loop on 150 Waymo Open Motion Dataset scenarios in Waymax against reactive and proposal-and-select baselines, and introduce a log-independent protocol that separates safety and regulatory compliance from resemblance to the recorded human trajectory. Under this protocol, W-SQP shows no systematic group-level deficit relative to expert replay on the log-independent safety and regulatory rules, with several localized regressions in the hardest, highest-divergence scenarios. The results characterize W-SQP as an auditable, priority-biased, anytime-capable NMPC prototype rather than a hard-real-time or formally safe controller.

Figures

Figures reproduced from arXiv: 2607.10975 by Benedikt Walter, Chaitanya Shinde, Hadi Hajieghrary, Miguel Hurtado, Paul Schmitt.

Figure 1
Figure 1. Figure 1: W-SQP closed-loop architecture. Each cycle, the rulebook and current scene are compiled into di [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: From priority-ordered rulebook to tiered shared slacks. The nine rule [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Sparsity pattern of the W-SQP NLP constraint Jacobian [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Empirical CDF of per-solve wall-clock time over the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Solver-time-limit sweep on a 40-scenario subset. As the limit decreases from unlimited to 30 ms, the log-independent safety/regulatory difference re￾mains near parity (blue, left axis), the log-coupled difference becomes more negative (orange), and the budget-hit rate increases smoothly (grey, right axis). 5.2. Compliance metrics as functionals Each rule r defines a violation predicate ϕr and an appli￾cabi… view at source ↗
Figure 7
Figure 7. Figure 7: places every planner in the plane of the two axes. Driving quality (horizontal) separates the planners by only ≈2 pp; imita￾tion (vertical) spreads them by ≈57 pp; and the expert sits on the y=100 dominance boundary of Prop. 1. A fair score reads the horizontal axis alone, so its iso-lines are vertical. The naive ag￾gregate’s iso-lines are instead tilted by the 16:2 group sizes, so it ranks a two-dimension… view at source ↗
Figure 8
Figure 8. Figure 8: Tier-separation mechanism: sweeping the priority ratio [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Auditability as a measured property. (a) Threshold-normalized per-surrogate residual raster for one scenario—every cell is a residual divided by its own threshold, so the single colorbar is meaningful across rows: the safety surrogates stay inactive while the low-priority route surrogate carries the late-scene relaxation. (b) Per-surrogate relaxation frequency across the audited fleet, with per-scenario sp… view at source ↗
Figure 10
Figure 10. Figure 10: Mean paired compliance difference (W-SQP − expert) by taxonomy group, bootstrap 95% CIs (n=150). The gap is confined to the two log-coupled rules; the log-independent group shows no systematic deficit and has a tight per￾scenario distribution around zero, whereas the log-coupled group is heavy-tailed. Individual rules move both ways ( [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 13
Figure 13. Figure 13: Log-coupled ∆c vs. exogenous interaction density at R=30 and 50 m: positive/flat slopes (the deficit does not grow with crowding). Inset: the circular solve-time proxy gives the opposite, misleading slope [PITH_FULL_IMAGE:figures/full_fig_p015_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: H6 — Stop-controlled. Expert vs. W-SQP rollouts at maximum divergence (left, centre) and the executed-vs-logged trajectory overlay (right). W-SQP takes a valid alternative line with near-parity safety (∆inv=−0.9 pp) and a −36.0 pp log-coupled deficit; the per-rule breakdown is in [PITH_FULL_IMAGE:figures/full_fig_p016_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: H6 — Rule-activity timeline (expert | W-SQP; light = applicable, solid = violated). The log-coupled rules (path tracking, route adherence) are elevated for W-SQP while the safety and regulatory rules stay close to expert replay: the −36.0 pp log-coupled deficit of [PITH_FULL_IMAGE:figures/full_fig_p016_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: H7 — Dense vehicle traffic (290 vehicles). Bird’s-eye keyframes (early / max-divergence / late) and the executed-vs-logged overlay. W-SQP accepts a different gap and is marginally safer (∆inv=+0.4 pp) while its log-coupled compliance drops −15.0 pp; the per-rule mechanism over time is in [PITH_FULL_IMAGE:figures/full_fig_p017_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Prop. 1 in time (H7, W-SQP solid vs. expert dashed; all quantities [PITH_FULL_IMAGE:figures/full_fig_p018_17.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

19 extracted references · 9 linked inside Pith

  1. [1]

    arXiv:2605.15120

    CLOVER: Closed-loop value estimation and ranking for end-to-end autonomous driving planning. arXiv:2605.15120. Benjamini, Y ., Hochberg, Y .,

    Pith/arXiv arXiv

  2. [2]

    8536–8542

    Liability, ethics, and culture-aware behavior specifica- tion using rulebooks, in: 2019 International Conference on Robotics and Automation (ICRA), pp. 8536–8542. Cheng, J., Chen, Y ., Chen, Q.,

    2019

  3. [3]

    arXiv:2404.14327

    Pluto: Pushing the limit of imitation learning-based planning for autonomous driving. arXiv:2404.14327. Cheng, J., Chen, Y ., Mei, X., Yang, B., Li, B., Liu, M.,

    Pith/arXiv arXiv

  4. [4]

    arXiv:2309.10443

    Rethinking imitation-based planner for autonomous driving. arXiv:2309.10443. Collin, A., Bilka, A., Pendleton, S., Tebbens, R.D.,

    Pith/arXiv arXiv

  5. [5]

    Safety of the in- tended driving behavior using rulebooks, in: 2020 IEEE Intelligent Vehicles Symposium (IV), pp. 136–143. Dauner, D., Hallgarten, M., Geiger, A., Chitta, K.,

    2020

  6. [6]

    ArXiv:2411.04653

    IGDrivSim: A benchmark for the imitation gap in autonomous driving, in: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). ArXiv:2411.04653. Gulino, C., Fu, J., Luo, W., Tucker, G., Bronstein, E., Lu, Y ., Harb, J., Pan, X., Wang, Y ., Chen, X., Co-Reyes, J.D., Agarwal, R., Roelofs, R., Lu, Y ., Montali, N., Mougin, P., Yang, Z., ...

    Pith/arXiv arXiv

  7. [7]

    arXiv preprint arXiv:2510.14677

    When planners meet reality: How learned, reactive traffic agents shift nuplan bench- marks. arXiv preprint arXiv:2510.14677 . Halder, P., Christ, F., Althoff, M.,

    arXiv

  8. [8]

    Hasuo, I.,

    Lexicographic mixed-integer motion planning with stl constraints, in: 2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC). Hasuo, I.,

    2023

  9. [9]

    arXiv preprint arXiv:2206.03418

    Responsibility-sensitive safety: an introduction with an eye to logical foundations and formalization. arXiv preprint arXiv:2206.03418 . Kanoun, O., Lamiraux, F., Wieber, P.B.,

    Pith/arXiv arXiv

  10. [10]

    UKACC International Conference on Control (Control 2000), Cambridge, UK

    Soft constraints and exact penalty func- tions in model predictive control, in: Proc. UKACC International Conference on Control (Control 2000), Cambridge, UK. Lai, L., Fiaschi, L., Cococcioni, M., Deb, K.,

    2000

  11. [11]

    Natural Computing 22, 227–242

    Pure and mixed lexicographic-paretian many-objective optimization: state of the art. Natural Computing 22, 227–242. doi:10.1007/s11047-022-09911-4. Lin, Y ., Xing, Z., Han, X., Althoff, M.,

    doi:10.1007/s11047-022-09911-4

  12. [12]

    Maierhofer, S., Rettinger, A.K., Mayer, E.C., Althoff, M.,

    Formalization of intersection traffic rules in temporal logic, in: 2022 IEEE Intelligent Vehicles Symposium (IV). Maierhofer, S., Rettinger, A.K., Mayer, E.C., Althoff, M.,

    2022

  13. [13]

    Manheim, D., Garrabrant, S.,

    Formalization of interstate traffic rules in temporal logic, in: 2020 IEEE Intelligent Vehicles Symposium (IV). Manheim, D., Garrabrant, S.,

    2020

  14. [14]

    arXiv preprint arXiv:1803.04585

    Categorizing variants of goodhart’s law. arXiv preprint arXiv:1803.04585 . Mavrotas, G.,

    Pith/arXiv arXiv

  15. [15]

    arXiv preprint arXiv:2511.10403

    nuplan-r: A closed-loop planning benchmark for autonomous driving via reactive multi-agent simulation. arXiv preprint arXiv:2511.10403 . Schwenzer, M., Ay, M., Bergs, T., Abel, D.,

    arXiv

  16. [16]

    arXiv preprint arXiv:1708.06374

    On a formal model of safe and scalable self-driving cars. arXiv preprint arXiv:1708.06374 . Tercan, A., et al.,

    Pith/arXiv arXiv

  17. [17]

    arXiv preprint arXiv:2408.13493

    Thresholded lexicographic ordered multiobjective optimization. arXiv preprint arXiv:2408.13493 . Wächter, A., Biegler, L.T.,

    Pith/arXiv arXiv

  18. [18]

    https: //github.com/waymo-research/waymax

    Waymax agents: IDM route-following policy. https: //github.com/waymo-research/waymax. Waymax documentation, ac- cessed 2026-02-13. Weihs, L., Jain, U., Liu, I.J., Salvador, J., Lazebnik, S., Kembhavi, A., Schwing, A.,

    2026

  19. [19]

    ArXiv:2007.12173

    Bridging the imitation gap by adaptive insubordina- tion, in: Advances in Neural Information Processing Systems (NeurIPS). ArXiv:2007.12173. Xiao, W., Mehdipour, N., Collin, A., Bin-Nun, A.Y ., Frazzoli, E., Tebbens, R.D., Belta, C.,

    Pith/arXiv arXiv 2007