- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
PhyDrive is a world model based framework for selecting among candidate driving trajectories. It keeps the pretrained DrivoR camera encoder and trajectory generator frozen, while a cmaera-LiDAR world model and Physical Evidence Guided Scorer (PEGS) evaluate 64 candidates using spatial, temporal, progress, and kinematic cues. A training only Planning-oriented Latent Representation Learning (PLRL) objective uses route conditioned auxiliary trajectories with safety factor supervision to encourage planning relevant world representatons. Ablations examine the contributions of PEGS, PLRL.
Critical strengths
The evaluation isolates trajectory selection from trajectory generation. PhyDrive keeps the same frozen set of 64 DrivoR proposals and improves performance over the matched DrivoR scorer, reaching 95.3 PDMS on NAVSIM v1 and 93.0 EPDMS on v2. This makes the reported gains meaningful evidence that the proposed scoring approach improves candidate ranking.
Critical weaknesses
-
The methodological contribution is not sufficiently clear or reproducible. PhyDrive combines physical-evidence features with trajectory ranking, but its distinction from DrivoR’s factor-wise trajectory scoring and prior world-model or safety-based trajectory-selection methods remains unclear.
-
The evaluation does not establish performance in interactive closed-loop driving. The paper also omits Navhard because its second stage does not provide LiDAR, leaving the method's performance under that evaluation setting untested.
Questions for authors
-
Would an evaluation of PhyDrive on Bench2Drive be feasible? Since Bench2Drive provides LiDAR and supports closed-loop evaluation, this would help determine whether the reported gains persist when the ego vehicle's actions affect subsequent observations. If the current model or setup is incompatible with Bench2Drive, what specific limitation prevents the evaluation?
-
Compared with DiffE2E, SafeDrive in Table 1, PhyDrive has a much higher EP score but lower NC and TTC scores. Does PhyDrive use DrivoR's 5-second long trajectory setting? Is the EP gain mainly due to selecting faster or farther reaching trajectories and what safety-progress trade off, if any, does this create?
-
What architecture is used for the world model and its occupancy and drivable area prediction heads?
3 comments
Dear Reviewer,
I hope this message finds you well. As the discussion period is nearing its end with less than two days remaining, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we're eager to address any remaining issues to improve our work.
Thank you for your time and effort in reviewing our paper.
Thank you for addressing my concerns. I will raise my rating to 3.
Clarifications and Additional Experimental Results
Thank you for your time and constructive feedback. In response to your comments, we have clarified the relevant descriptions and added experimental results to address your concerns. We will incorporate these changes into the main manuscript and supplementary material.
Q1:The evaluation of PhyDrive on Bench2Drive.
A1: We added a closed-loop evaluation on Bench2Drive. As shown below table, compared with the baseline, PhyDrive retains its performance advantage, suggesting its gains persist when ego actions influence subsequent observations. In closed-loop evaluation, higher route completion and driving efficiency alongside a lower collision rate suggest that physical-evidence-guided trajectory selection balances safety and progress.
Q2.1:Does PhyDrive use DrivoR's 5-second long trajectory setting?
A2.1: No, PhyDrive uses a 4-second prediction horizon. To examine the effects of the long-trajectory setting on PhyDrive, we additionally evaluate a 5-second setting, with results shown below. Longer-horizon trajectory supervision provides only limited additional progress gains for PhyDrive, suggesting that physical-evidence guidance already supports effective progress optimization with limited dependence on the long-trajectory setting.
Q2.2:Is the EP gain mainly due to selecting faster or farther reaching trajectories and what safety-progress trade off, if any, does this create?
A2.2: During training, we consistently supervise all factors with equal weights, without assigning additional loss weight to either safety or efficiency. The physical-evidence ablation shows improvements across all evaluated factors, indicating that the gains reflect an overall improvement in trajectory selection without an explicit preference for faster or farther-reaching trajectories.
To further explore the safety–progress trade-off, we tested adjustments to the EP weight in the inference-time score aggregation formula. These results show that the preference between safety and progress can be adjusted through inference-time aggregation weights without retraining the model.
Q3: What architecture is used for the world model and its occupancy and drivable area prediction heads?
A3: Inspired by WoTE [1] and ResWorld [2], we construct a lightweight world model. We first fuse per-frame point-cloud and image features, then use a temporal module to extract information across frames. Finally, a residual recurrent network predicts the current and future states.
Following WoTE [1], we use lightweight convolutional prediction heads. Our implementation shares a feature transformation module and uses separate output layers to predict future occupancy maps and the current drivable-area map, respectively. The occupancy head shares parameters across future timesteps.
W1. The methodological contribution is not sufficiently clear or reproducible.
A: To distinguish similar trajectories with different safety outcomes, we construct candidate-specific, factor-aligned physical evidence to guide interactions between trajectories and future-scene representations. We commit to releasing the complete code to support reproducibility.
Comparison with DrivoR. DrivoR uses factor-wise scoring, with physical cues implicit in learned features. We construct explicit evidence through footprint, spatial-margin, and temporal-lookahead queries to guide factor-specific encoding and scoring.
Comparison with prior methods. WoTE [2] evaluates trajectories using future BEV, while SafeDrive [3] predicts safety through sparse agent interactions. We query dense predicted maps along candidate vehicle footprints over time, organizing occupancy and road-boundary relationships into factor-aligned evidence that guides scoring.
[1] End-to-End Driving with Online Trajectory Evaluation via BEV World Model. ICCV 2025.
[2] ResWorld: Temporal Residual World Model for End-to-End Autonomous Driving. ICLR 2026.
[3] SafeDrive: Fine-Grained Safety Reasoning for End-to-End Driving in a Sparse World. CVPR 2026.