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