acceptodds
Under review as a conference paper at ICLR 2027

PHLEX: Physically Layered Explanation for Counterfactual Motion Planning

Abstract

Predicted trajectories describe where an agent may move, but provide limited insight into the physical mechanisms behind that motion and are difficult to systematically vary for counterfactual reasoning. We introduce PHLEX (PHysically Layered EXplanation), a predictor-agnostic framework that maps predicted motion into a structured physical mechanism representation. PHLEX progressively explains each trajectory through kinematic models, longitudinal and lateral control laws, and scene references. Together, these layers form a structured physical explanation of how the predicted motion is generated and grounded in the scene. These explanations can be executed and varied to generate physically grounded counterfactual futures, which are further used to guide ego planning. This connects the flexibility of learned trajectory prediction with explicit physical interpretation and controllable what-if reasoning. Experiments across multiple predictors and datasets show that PHLEX improves physical feasibility and planning safety while preserving trajectory accuracy.

Primary area:
applications to robotics, autonomy, planning
Keywords:
autonomous driving, motion prediction, motion planning, physical interpretability, counterfactual reasoning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_sV4HOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PHLEX represents predicted trajectories as combinations of kinematic models, longitudinal and lateral control laws, and scene references. It uses fitting and reconstruction errors to retain multiple compatible explanations, then executes these explanations to generate alternative futures. A learned Refiner uses the resulting trajectories, mechanism evidence, and interaction risks to predict a residual correction to the host ego plan. Experiments with several frozen host models on nuScenes and Argoverse 2 show improvements in physical quality and open-loop collision rates while largely preserving trajectory accuracy. The supplementary material describes candidate construction, continuous execution across segments, the Refiner architecture, and training.

Critical strengths
  1. PHLEX represents predicted trajectories using executable combinations of kinematic models, control laws, and scene references, allowing multiple compatible explanations for the same trajectory. The supplementary material describes how constraints are enforced at each step, how state continuity is maintained across segments, and how missing evidence is handled. The paper also distinguishes compatibility with a predicted trajectory from evidence of the driver’s actual intent.

  2. PHLEX improves the physical quality of trajectories across several frozen host models and two datasets, while largely preserving prediction accuracy. It also reduces open-loop collision rates for all three planning hosts. Ablation results show that both the mechanism library and rollout-risk features contribute to these gains.

Critical weaknesses
  1. The contribution of the layered representation to reconstruction performance is not sufficiently explained, and the procedures for selecting explanations and constructing forecasting outputs remain unclear, limiting the interpretability and reproducibility of the results. (Q1–Q3)

  2. The baseline comparisons do not sufficiently establish whether the improvements arise from the proposed layered representation rather than generic neural correction or simpler physical constraint enforcement. (Q4)

  3. The experiments provide limited direct evidence of the method's counterfactual capability, particularly regarding the effects of mechanism interventions and the modeling of interactions with surrounding agents. (Q5)

Questions for authors
  1. What accounts for the reconstruction-error gap between layered and flat representations? How does the change from layered to flat representation affect candidate construction, support weighting, and explanation selection? What accounts for the reconstruction-error gap in Table 4 when the mechanism library and candidate budget are held fixed?

  2. Which explanation weights are used for reconstruction evaluation? In Appendix D.1, which weights are used to select the highest-support explanation for the reconstruction metric: the original physical weights or the temporally refined weights?

  3. How are the final multimodal forecasts constructed? For Table 3, how is a single forecast selected or constructed from the explanation sequences for each host mode? Are fallback trajectories included in all reported forecasting metrics?

  4. Are the baseline comparisons sufficient and fair? Does the neural-only corrector in Table 2 have comparable model capacity and use the same training objective as PHLEX? Please specify its architecture, parameter count, inputs, and loss terms. How does PHLEX compare with a physical projection or constrained fitting baseline?

  5. How are counterfactual interventions and interactions evaluated? When constructing the cached surrounding-agent futures in Appendix B.2, which ego trajectory is used as a reference for scene-dependent control laws? Can these futures capture responses to different ego candidates? For the intervention example in Section 6.4, could the authors show how the resulting trajectories and interaction risks affect the final ego plan? In Table 5, does removing rollout-risk features leave both safety-loss terms in Eq. (16) unchanged?

Reviewer_wkdeOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper is concerned with learned ego for autonomous driving. The idea is to score multiple predictions for the ego plan and other agent future motions against physical plausibility with handcrafted logic, and then feed these plausibility scores to a “refiner” network that picks that ego plan from candidate predictions and plausibility scores. Plausibility scores involve match with different vehicle dynamics models; match with scripted planner models (like IDM, follow-the-leader, etc), and relation to map references (stop lines, etc). The method applies to end-to-end planners, but to compute scores the methods assumes access to HD maps or explicit map predictions.

Critical strengths
  1. The idea is interesting and to my knowledge novel.
  2. The method makes sense on an intuitive level (after careful reading)
  3. Presented results show advantage for the method over multiple planner network families
  4. The paper is mostly well written, with the exceptions mentioned below.
Critical weaknesses
  1. Phlex seems to require access to HD map information or map predictions (for stop-lines, IDM, etc) that is not typically assumed to be available to end-to-end planning networks. It is unclear what method is used in the paper to access / predict the required map information, and inaccurate online map predictions may impact the overall efficacy. The paper should be upfront about this issue and provide some evaluation against partial inaccurate maps.
  2. Open-loop only evaluation. Providing proxy closed-loop evaluation results would largely increase the impact, such as from NAVSIM or AlpaSim.
  3. I found it very hard to understand the context that this paper is ultimately about ego planning for autonomous driving as a) abstract doesn’t mention anything about the fact that we are in the autonomous driving domain b) the intro first paragraph positions the paper wrt. motion forecasting. It took me a long time to figure out the phlex explanations are actually applied to both ego plans and agent predictions.

I believe a rebuttal that addresses these concerns would make this paper a weak accept.

--- AFTER Rebuttal Thank you for the careful response and new evidence. My concerns are addressed, im raising my score to weak accept.

Questions for authors

none

1 comment

Thank you for the careful response and new evidence. My concerns are addressed, im raising my score to weak accept.

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

This paper proposes PHLEX, a predictor-agnostic framework that converts predicted trajectories into structured physical explanations. It decomposes motion into three layers: kinematic models, control laws, and scene references. Multiple compatible explanations are retained and executed to generate counterfactual futures, which are then used to refine ego planning through a lightweight neural network. Experiments on nuScenes and Argoverse 2 across multiple predictors demonstrate improvements in physical feasibility, planning safety, and trajectory quality.

Critical strengths
  • ** Interpretable and executable motion representation.** The hierarchical decomposition into physical mechanisms provides a structured way to explain predicted motion and generate controllable counterfactual futures, going beyond conventional trajectory prediction.

  • Consistent improvements across different predictors. PHLEX significantly reduces kinematic violations and collision rates across multiple host models. The ablation studies further support the effectiveness of the layered representation and counterfactual risk features.

Critical weaknesses
  • Insufficient evidence for practical planning benefits. All evaluations are open-loop, and comparisons with stronger physics-constrained trajectory refinement or safety optimization baselines are missing. Therefore, it is difficult to isolate the benefits of mechanism-level counterfactual reasoning.
Questions for authors
  • Q1: How does PHLEX compare with physics-constrained trajectory smoothing or collision-aware refinement methods under comparable settings?

  • Q2: Can the authors provide closed-loop planning results and report the computational overhead of counterfactual generation?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.