- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper introduces CoReDrive, an agent-conditioned autonomous driving planner that employs a counterfactual response supervision mechanism. The method characterizes agents by leveraging detected geometric information, multimodal future trajectories, and motion-inference-based intentions.
During training, the system constructs counterfactual interaction scenarios—without requiring additional sensor rendering—by modifying the predicted future trajectory of a specific agent while keeping current observations and scene geometry intact. Since the modified scenarios lack "expert-level" ego-vehicle trajectories, the system utilizes a "geometric teacher" mechanism to generate a set of feasible ego-vehicle responses rather than a single pseudo-expert trajectory.
This set of responses guides the filtering and final ranking of candidate trajectories; simultaneously, paired factual and counterfactual geometric risk objectives are used to train an interaction-scoring calibration model, which plays a role in velocity feasibility verification and final trajectory selection. The method has been evaluated on the NAVSIM platform.
Critical strengths
-
The approach to constructing a “response set” is sound and well-founded. Since edited interactive scenarios typically lack a single “correct” response for the ego vehicle, constructing a set of spatiotemporally feasible trajectories is more reasonable than specifying a single synthetic expert trajectory.
-
The counterfactual supervision mechanism is computationally practical. Directly editing the predicted future trajectories of actors at the planning interface level avoids the high costs associated with sensor-level re-rendering or full-scene simulation.
-
Qualitative examples clearly demonstrate conditional behaviors. The paper illustrates how actor information influences ego vehicle planning, showing that changes—such as a lead vehicle braking or a roadside vehicle pulling into the lane—elicit distinct responses from the ego vehicle.
Critical weaknesses
-
Core Contribution—The quantitative gains from "counterfactual response learning" are negligible. Under the matched ResNet-34 setting, the baseline model achieves an EPDMS of 90.1; introducing "actor modeling/conditioning" raises this to 90.7; however, adding the proposed "counterfactual response learning" results in a marginal increase from 90.7 to only 90.8. Consequently, the majority of the performance improvement relative to the baseline appears to stem from "actor conditioning" rather than the "counterfactual learning" highlighted in the paper. Due to the absence of multiple random seed tests, confidence interval analysis, or statistical significance testing, it is difficult to determine whether this 0.1-point gain is robust.
-
Lack of systematic quantitative evaluation regarding core "counterfactual capabilities." The method defines various intervention types—such as lead vehicle braking, crossing/cutting in, turning across paths, pulling out from the roadside, obstructed lane changes, and failure to yield. However, the paper reports neither response success rates, collision avoidance performance, nor the recall of feasible responses; nor does it analyze performance variations based on intervention type or severity. Evidence regarding counterfactual capabilities relies primarily on cherry-picked, open-loop qualitative examples, which is insufficient to demonstrate that the proposed training method consistently improves behavioral performance across diverse intervention scenarios.
-
The counterfactual supervision mechanism remains heavily reliant on manually designed rules. The intervention library and the "geometric teacher" depend on human-defined response requirements, temporal constraints, collision detection, safety buffer margins, kinematic thresholds, risk functions, and quality metrics based on progress assessment. This raises the question: has the model truly learned generalized counterfactual interaction reasoning, or has it merely distilled a set of predefined driving heuristics? Furthermore, the method fails to demonstrate generalization capabilities regarding unseen interaction types.
-
The "response teacher" is constrained by the existing candidate trajectory support mechanism. ** The formulation of counterfactual interactions remains unidirectional and may deviate from the data manifold. ** The method modifies the future trajectories and intentions predicted for the actor while keeping camera observations and the current geometric state unchanged. Although these modifications are filtered for kinematic plausibility, this does not guarantee consistency with visual cues or the full joint scene distribution. Furthermore, other road users do not react to the ego vehicle's altered behavior. This limits the method's applicability to scenarios involving bidirectional interactions, such as game-theoretic situations, merging, and mutual yielding.
Questions for authors
-
Is the improvement in the EPDMS metric from 90.7 to 90.8 statistically significant across multiple training seeds? If so, please provide the mean and standard deviation.
-
Could the authors provide a quantitative evaluation comparing the "Actor-only" model and the "Actor + Response" model specifically on the generated counterfactual scenarios (rather than primarily on the factual NAVSIM benchmark)?
-
Could the authors report the "response success rate" and "collision/conflict avoidance performance" separately for each intervention family and intervention severity level?
-
What proportion of the adopted counterfactual scenarios contains at least one feasible response that lies within the retained candidate-path support set?
-
How does the method handle counterfactual interactions where an appropriate ego response requires a path that was discarded during the initial path-pruning stage?