acceptodds
Withdrawn from ICLR 2027

Learning to Explain While Planning: Rule-Aligned Diffusion Planning for Autonomous Driving

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Diffusion planners exhibit strong capabilities in generating multimodal trajectories. However, existing methods primarily rely on expert demonstrations to fit trajectory distributions, learning statistical correlations among scenes, behaviors, and trajectories without explicitly modeling driving rules. In long-tail scenarios where expert data are scarce, the lack of behaviors to imitate may lead to trajectories that violate safety or compliance requirements. Moreover, their generation process lacks rule-level explanations, making it difficult to determine which rules drive trajectory adjustments, when they take effect, and how strongly they act, thereby limiting failure diagnosis, safety validation, and targeted improvement. To address these limitations, we propose the Rule-Aligned Diffusion Planner (RADP), which incorporates differentiable driving rules into the diffusion objective during training, turning rule knowledge into intrinsic behavioral principles beyond finite demonstrations. We further introduce Rule-Pressure Attribution (RPA), which constructs supervision signals from gradients of rule losses with respect to predicted trajectories and employs a lightweight attribution head to estimate the optimization pressure exerted by each rule online. To assess the closed-loop behavioral relevance of these attributions, we propose a temporal risk-alignment protocol that evaluates whether current rule pressures reflect corresponding risks during subsequent closed-loop execution. Experiments on nuPlan show that RADP improves closed-loop planning in challenging safety-critical scenarios, while RPA exhibits consistent temporal alignment with subsequent rule-specific risks, validating both intrinsic rule learning and rule-level interpretability.

Primary area:
generative models
Keywords:
Diffusion Models, Autonomous Driving, Motion Planning, Rule-Aligned Learning, Differentiable Constraints, Interpretable Planning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_stw8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces RADP, which adapts a pretrained diffusion planner using differentiable driving-rule penalties, and RPA, which distills the magnitudes of rule-cost gradients into an auxiliary prediction head. The head provides online rule-pressure estimates without modifying the generated trajectory. Evaluation covers nuPlan closed-loop planning, head–teacher agreement, and alignment with subsequent driving risks.

Critical strengths
  • The reactive closed-loop results show improvements in collision and TTC metrics, particularly on test14-hard. The rule-free LoRA ablation helps distinguish the contribution of rule supervision from additional fine-tuning alone.
  • The attribution evaluation goes beyond qualitative visualizations by including head–teacher comparisons, temporal controls, and scenario-level confidence intervals. These provide useful evidence of risk association, although not necessarily explanatory faithfulness.
Critical weaknesses
  • The methodological novelty appears limited. Planner adaptation uses differentiable penalties on predicted clean trajectories and the attribution module regresses calibrated gradient norms. Both are broadly applicable techniques, and the paper does not sufficiently establish what is technically new beyond their combination. The contribution specific to diffusion modeling is particularly unclear.
  • The student provides auxiliary outputs without affecting trajectory generation or planning decisions. Its value therefore depends on what it adds as an analysis tool, but this is not clearly established. Comparisons with simpler alternatives, such as direct rule costs or basic risk indicators, are needed to show whether the learned head offers more useful explanations.
Questions for authors
  • What distinguishes the planner objective from standard differentiable output regularization? Is there a diffusion-specific formulation or property beyond applying rule penalties to the predicted clean trajectory?
  • Since the student does not affect planning decisions, is its intended contribution primarily computationally efficient sensitivity estimation? What evidence shows that its outputs provide explanatory value beyond directly evaluating rule costs or simple risk indicators?
  • How does the rule regularization account for the noise level of XtX_t and the gap between the predicted conditional mean X^0(Xt,t,S)\hat{X}_0(X_t,t,S) and the final generated trajectory?
  • RADP improves collision and TTC metrics on test14-hard, but also reduces progress. Can you clarify whether these gains reflect better interaction handling or primarily more conservative driving? A comparison at matched progress levels would help distinguish the two.
Reviewer_xm6vOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RADP, which adapts a pretrained Diffusion Planner using six differentiable driving-rule losses. It also introduces an RPA head that approximates calibrated rule-gradient magnitudes to provide online rule-pressure estimates without computing exact gradients. Experiments on nuPlan evaluate closed-loop planning, head–teacher agreement, future-risk alignment, and computational overhead, showing improved safety metrics on the reactive test splits.

Critical strengths
  1. Integrating explicit rule priors into learned planning addresses a valuable problem, particularly when expert demonstrations provide insufficient coverage of safety-critical situations.

  2. The reported reactive test results show improvements in planning scores and safety metrics, especially on test14-hard. RADP also substantially outperforms the reported collision-only inference-time guidance baseline.

Critical weaknesses
  1. The planning contribution is insufficiently distinguished from existing auxiliary-loss approaches. Prior work on learning-based planning and end-to-end autonomous driving, such as PLUTO [1] and VAD [2], already incorporates differentiable safety or compliance constraints during training. The incremental advance beyond applying such losses to a diffusion planner remains unclear.

  2. The rule objectives involve numerous weights, scales, gates, and aggregation choices, but the ablations provide limited justification for this complexity. Sensitivity analyses and per-rule ablations are missing, leaving robustness and the safety–progress trade-off insufficiently understood.

  3. The independent value of RPA remains under-validated. The head does not affect planning, while agreement with a gradient teacher and correlation with future risks do not establish what explanatory information it provides beyond direct rule costs or violation measures.

[1] Cheng et al. PLUTO: Pushing the Limit of Imitation Learning-based Planning for Autonomous Driving. 2024. [2] Jiang et al. VAD: Vectorized Scene Representation for Efficient Autonomous Driving. ICCV 2023.

Questions for authors
  1. What distinguishes RADP from conventional auxiliary-loss training beyond the choice of a diffusion backbone and a broader rule set? Can a controlled comparison against simple collision and drivable-area losses establish the incremental benefit?

  2. How robust are the results to the rule weights, penalty shape, physical scales, and normalization choices? Which components are necessary, and which rules account for the safety gains and reduced progress?

  3. What is the intended practical role of the RPA head: efficient monitoring, explaining planning decisions, or another use? What additional value does it provide over directly reporting rule costs or violation magnitudes?

Reviewer_DH4UOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The submission fine-tunes Diffusion Planner on nuPlan. It adds six hand-designed differentiable rule costs to the x0x_0-prediction loss: collision, lane boundary, overspeed, kinematic limits, comfort and goal progress. Only LoRA adapters on the decoder FFNs are trained.

On top of the frozen fine-tuned planner, the authors define a per-rule “pressure”: the log-calibrated RMS norm of ∇XJi\nabla_X J_i at the generated trajectory. They distill this into a small attention + MLP head (RPA) that outputs six non-negative scores alongside each plan.

The attributions are validated with a “temporal risk-alignment” protocol. Each channel at planning frame tkt_k is compared against a rule-matched severity measured on the executed rollout over [tk,tk+Hi][t_k, t_k+H_i]. The window is 2 s for collision/TTC and 8 s otherwise. The metrics are within-scenario Spearman, AUPRC and Lift@10%, with a within-scenario shuffle as the control.

Critical strengths
  1. The rules and the evaluation endpoints are specified concretely enough to reimplement (Appendix B, Table 8). The risk endpoints come from executed rollout states, not from the training costs. Windows, thresholds and aggregation operators are stated to be fixed before inspecting results. There is a scenario-level bootstrap and a shuffle control. Most interpretability work in driving stops at attention maps and a few hand-picked figures, so this is more care than usual. It’s the main reason I think there is something here worth building on, even though I don’t think the protocol shows what the paper says it shows (W2).

  2. The training-time rule loss is cheap to adopt: LoRA on twelve linear layers and no change at inference. On test14-hard-r, collision goes from 86.95 to 92.10 and TTC from 79.78 to 83.46. The no-rule LoRA control (88.42/81.62) suggests part of that comes from the rules rather than from extra fine-tuning. If this holds across seeds and splits, it is a useful practical result.

Critical weaknesses
  1. The planning gains are small, mixed, and not separated from noise or simple conservatism:
  • The headline result is a ~1.7-point reactive gain on test14-random and test14-hard. It comes from a single seed (3407), with no variance reported, on splits of a few hundred scenarios.
  • Reactive val14 drops by 1.3. Non-reactive is worse on two of three splits, including test14-hard (75.01 → 74.80), which is the split the safety claim leans on.
  • Progress drops on every split (92.65 → 90.07 on test14-hard). “Fewer collisions, less progress” is also what a more cautious planner would give you, and nothing in the paper rules that out.
  • The gain attributable to the rules rather than to extra LoRA training is 70.62 vs 69.20, on one split. That same split is used for the ablation, latency, qualitative and Appendix D results, and the paper doesn’t say where λrule\lambda_{rule} and wiw_i were selected.
  • The comparison in Table 7 doesn’t support the training-vs-inference argument. Collision-guided sampling drops the collision metric from 86.95 to 76.65, i.e. it collides more than the unguided planner, and comfort falls to 60.66. That looks like a mis-tuned baseline, not evidence that training-time integration is better.
  • Test14-hard is selected by planner difficulty, not data scarcity, so the long-tail claim in the abstract isn’t tested.
  • “Highest reported scores” (Table 3) only holds after excluding rule-based and hybrid planners such as PDM-Closed and the refinement variants reported in the PLUTO and Diffusion Planner papers.
  • Training-time rule losses on top of imitation are not new; ChauffeurNet already used collision and on-road losses.
  1. The attribution evaluation doesn’t test the explanatory claim:
  • The abstract promises to identify “which rules drive trajectory adjustments, when they take effect, and how strongly they act.” What is computed is the gradient norm of hand-written costs with respect to the output of a frozen planner. At inference the rules play no role in producing the trajectory, so this cannot tell us what drove it.
  • It is a post-hoc evaluation of the output under an external cost model, which is the category the introduction criticizes. It also doesn’t depend on RADP training: the same teacher can be computed on the unmodified Diffusion Planner.
  • Figure 3 shows the gap. Lane pressure climbs after frame 80 while corridor clearance stays negative for most of the remaining rollout. The head is reporting a violation the planner never corrects.
  • The teacher is close to a function of the cost itself. With φ≈[max⁡(0,z/σ)]2\varphi \approx [\max(0, z/\sigma)]^2 and roughly unit-norm ∇Xz\nabla_X z, ∥∇XJi∥\|\nabla_X J_i\| is close to a monotone function of JiJ_i (roughly ∝Ji\propto \sqrt{J_i}, up to the number of valid elements). Every alignment metric is rank-based, so I’d expect JiJ_i to produce nearly identical numbers. Section 5 argues the gradient is the better target but never compares the two.
  • The protocol measures plan–execution consistency. With 10 Hz replanning, the executed trajectory follows the plan closely. So violations in the plan at tkt_k correlating with violations in the next 2 to 8 s of execution is expected by construction.
  • The head’s efficiency argument is weak. The teacher only differentiates analytic costs with respect to an H×dH\times d tensor; it does not backpropagate through the network. The exact teacher on cached states already costs 62.5 ms, and is unoptimized. The head approximates it at Spearman 0.55 and 71% Top-1, with no majority-class reference for Top-1. Giving up exactness for an approximation that loose needs a better justification.
Questions for authors
  1. How many seeds were run, and what is the variance per split? Which split was used to choose λrule\lambda_{rule}, wiw_i and the checkpoint? If it was test14-hard, the test14-hard numbers in Tables 2, 3 and 7 are not held out.

  2. For the collision-guided baseline: what guidance scale and steps were used, and was it tuned? Why does it increase collisions? Using the guidance setup released with Diffusion Planner would be the fair comparison.

  3. Which neighbor futures enter the collision cost? Specifically, those used during training (logged ground truth or the model’s own joint predictions), and those used when computing the teacher in closed loop.

  4. For each endpoint, how many scenarios are eligible, and what is their mean per-scenario positive rate (the correct AUPRC/Lift reference)?

  5. Why does the online exact teacher add ~214 ms when the cached computation takes 62.5 ms?

  6. For Appendix D: on test14-hard-r, in how many scenarios does RADP collide where the baseline does not?

  7. Risk-alignment numbers for three baselines: persistence (risk at tkt_k), raw JiJ_i on the planned trajectory, and min clearance/TTC at tkt_k.

  8. The teacher and head computed on the unmodified Diffusion Planner. If alignment is similar, RADP and RPA are independent contributions, and rule-aligned training doesn’t make the explanations more meaningful.

  9. Even a small intervention study would help, for example rescaling one wiw_i and checking whether the attribution shift predicts the change in trajectory or risk. This is the experiment that would test “which rules drive adjustments.” Section 5 lists it as future work, but the central claim depends on it.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.