SAH-Drive: A Scenario-Aware Hybrid Planner for Closed-Loop Vehicle Trajectory Generation

We really appreciate the reviewer’s constructive comments and positive feedback, which have helped us better articulate our contributions and clarify the novelty of our approach. Regarding the concerns of the reviewer 2ndh, we provide the following responses.

Q1: Even though SAH-Drive has a scenario-aware switching logic, which is different from the existing hybrid planner approaches, the novelty is not strong.

Thanks for your valuable comments. The novelty of this work lies not only in proposing the Scenario-Aware Hybrid Planner Paradigm but also in the following key aspects:

• STDP-Based Decision Neuron: It memorizes the performance of both planners during the planning process, enabling adaptive switching based on scenario characteristics.
• Proposal Number Regulator: It dynamically adjusts the number of diffusion model samples in real-time, improving planning efficiency without sacrificing trajectory diversity.
• Trajectory Fusion Mechanism: We propose an exponential-weighted fusion of the highest-scoring diffusion and PDM trajectories to balance trajectory diversity and driving safety.
• Efficient Learning-Based Planner Adaptation: SAH-Drive reduces the required training data compared to fully learning-based planners (from 285.61GB to 8.25GB, a reduction of approximately 97%) while maintaining planning performance, especially in long-tail scenarios such as interPlan and Test14-hard (see rebuttal table 2 https://anonymous.4open.science/r/SAH-Drive-Pic-9E34/benchmark_table.png for details). This makes targeted training on rare, real-world long-tail scenarios both feasible and efficient.

We believe that adapting learning-based planners to long-tail scenarios with limited data is a promising direction for future research, and our work lays the foundation for this advancement. We will further highlight the novelty in the final version.

Once again, we sincerely appreciate your valuable insights, and we hope this clarification strengthens the perceived novelty and contributions of our work.

We sincerely appreciate the reviewer’s constructive feedback. In response to the concerns raised by Reviewer Z9oQ, we provide the following responses.

Q1: Detailed explanation on the proposal regulator.

The number of diffusion trajectories $N'$ is dynamically adjusted based on the highest diffusion trajectory score $S_{\max}$ relative to a threshold $\tau$.

The trajectory count is updated based on the highest trajectory score:

\begin{cases} \frac{N}{2}, & S_{\max} > \tau, \\ 2N, & S_{\max} < \tau. \end{cases}$$ To regulate trajectory count, we apply: $$N' = \max(N_{\min}, \min(N', N_{\max})).$$ We will add the detailed explanation of the proposal number regulator in the final version. > Q2: Including an algorithm to explicitly illustrate the switching workflow We outline the decision process (Fig. 3, p5 l220 in the original paper) and will add a detailed algorithm in the revised paper. **Input**: Rule-based planner score $s_c$, Learning-based planner score $s_r$ **Output**: Selected planner 1. Update $w_r$ and $w_c$ using Equation (3) from the paper. 2. Update consecutive counts $n_e$​ and $n_p$. 3. assign the decision category: category $\leftarrow$ decision_space$(s_r, s_c,n_e,n_p)$ 4. Select the planner based on the decision category: - If category = score, then planner $\leftarrow$ score_rule$(s_r, s_c)$ - If category = scenario, then planner $\leftarrow$ scenario_rule$(n_e, n_p)$ - Otherwise, planner $\leftarrow$ STDP_neuron$(w_r, w_c)$ > Q3: Lack of comparison with recent SOTA methods and further comparison on Test14-hard. We have extended our benchmark evaluation to include **recent learning-based SOTA methods**: Pluto, PlanTF, and DiffusionPlanner. The results are summarized in the table below: | Planner | interPlan | Val14 (R) | Val14 (NR) | test14 (R) | test14 (NR) | test14-hard (R) | test14-hard (NR) | | ----------------- | --------- | --------- | ---------- | ------------ | ----------- | --------------- | ---------------- | | Pluto | 48 | 78 | 89 | 78 | **89** | 60 | 70 | | SAH (PDM+Pluto) | 49(2%↑) | 79(1%↑) | 89 | **88**(13%↑) | 88 | 81(35%↑) | **78**(11%↑) | | PlanTF | 33 | 77 | 84 | 80 | 85 | 61 | 69 | | SAH (PDM+PlanTF) | 40(21%↑) | 78(1%↑) | 86(2%↑) | 86(8%↑) | 87(2%↑) | 80(31%↑) | 77(12%↑) | | Diffusion Planner | 24 | 83 | **90** | 83 | **89** | 69 | 75 | | SAH-Drive | **64** | **90** | 89 | 87 | 86 | **83** | **78** | As shown in the table, SAH-Drive consistently outperforms SOTA learning-based planners across multiple evaluation metrics. Notably, despite being trained only on nuPlan-mini, SAH-Drive achieves **SOTA scores on interPlan and test14-hard**, while also delivering **near-SOTA performance on other benchmarks**. The full experimental results, including all baselines discussed in the paper, can be found at https://anonymous.4open.science/r/SAH-Drive-Pic-9E34/benchmark_table.png. > Q4: Further experimental validation is needed by integrating PDM with different types of planners such as SAH + PlanTF and Pluto. Thanks for pointing this out. We employed the SAH + PlanTF and Pluto as the reviewer requested. The results are summarized in the table under Q3. As shown in the table, both PlanTF and Pluto achieve **consistent improvements across various nuPlan splits** when integrated with PDM under the SAH framework, outperforming their original versions. The performance gains are particularly notable on Test14-hard (R), with improvements of **35% and 31%**, respectively. > Q5: Extra results under different time variance between rule and learning-based planner for further justifications. To complement our analysis of time variance between rule-based and learning-based planners, we conducted an additional experiment using SAH-Drive in the regular straight-through intersection nuPlan scenario. The visualization is available at https://anonymous.4open.science/r/SAH-Drive-Pic-9E34/time_variance_visualization.png. The learning-based planner (yellow) was active **6.70%** overall. In contrast, its activation rate **increased** to **43.33%** in the long-tail scenario, highlighting the scenario-aware nature of our method. We will include the above details in the revised paper. > Q6: Insufficient references Thank you for your suggestion. We will incorporate **Pluto, PlanTF, Diffusion Planner, CarPlanner, and BeTop** into the related work section to provide a more comprehensive discussion of prior research.

We thank the reviewer for the constructive comments. Regarding the concerns of the reviewer NL4n, we provide the following responses.

Q1: Achieving SOTA only on the interPlan may not be enough, especially given that it fails to outperform the ensemble methods this paper used on Val14.

Thanks for the constructive comment. We understand why the reviewer might perceive a "1+1 < 2" issue—where integrating PDM-Closed and Diffusion-ES within the SAH paradigm appears to yield a lower score. However, this is a misunderstanding. The learning-based planner used in SAH-Drive is not the original Diffusion-ES, but rather a simplified version that removes evolutionary search and reduces the model complexity.

We simplify the Diffusion-ES to improve planning efficiency, achieving a nearly 98% reduction in average per-frame runtime—from 5s to just 0.1s. Additionally, our optimization drastically reduces the model’s memory footprint (from 25.62MB to 1.08MB, a 96% decrease) and parameter count (from 6.7M to 284K, also a 96% reduction). This lightweight design enables the model to achieve strong performance even when trained solely on nuPlan-mini. However, this efficiency gain comes with a slight performance trade-off.

As shown in the table below, when evaluated on Val14, this simplified diffusion-based planner score is 88. Additionally, SAH-Drive’s rule-based planner (PDM-Closed) operates at 5Hz, and its standalone performance on Val14 is 90. When combining these two planners within the SAH paradigm, the overall performance remains 90, which does not indicate a drop in performance compared to its individual components. This directly addresses the reviewer’s concern.

Q2: Score-based STDP is interesting, but my concern is that it isn’t effective in val14, raising questions about its generalization ability.

Thanks for pointing this out. Regarding generalization concerns, we conducted further experiments on two extra datasets: Test14-Random and Test14-hard. Results on interPlan, Val14, Test14-random, and Test14-hard are shown in rebuttal table 2 (https://anonymous.4open.science/r/SAH-Drive-Pic-9E34/benchmark_table.png). SAH-Drive achieves SOTA on interPlan and Test14-hard and near-SOTA on Val14 and Test14-random, which confirms SAH-Drive’s robustness and generalizability across diverse scenarios.

Q3: Most aspects aside from STDP feel somewhat unremarkable and the Scenario-based and Score-based switching rules seem somewhat naive.

We appreciate the reviewer for the comment but disagree with our highest respect that most aspects aside from STDP are unremarkable. The Scenario-based and Score-based switching rules play a crucial role in complementing the STDP-based decision neuron.

The Scenario-based and Score-based rules provide immediate decision-making capabilities, analogous to Short-Term Plasticity in biological neurons, which enables rapid adaptation to sudden environmental changes. If we were to rely solely on the STDP-based decision neuron, its memory-driven nature—which accounts for historical planning performance—could lead to delayed responses to sudden, strong signals. This delay in planner switching could be problematic, especially in autonomous driving scenarios where immediate adaptation is critical. The ablation experiment table in the original paper (p7, l350) further validates this, showing that without switching rules, the planner’s interPlan score decreases by 3.13%. Unlike approaches that emphasize model complexity and large-scale datasets, SAH-Drive reduces reliance on extensive training data through flexible planner switching.

Q4: The paper still relies on the PDM scorer to evaluate planner-generated trajectories, which often requires precise lane markings and surrounding vehicle motion information. Such information is often expensive to obtain in real-world scenarios.

Thanks for your valuable comments, and we understand your concerns. However, our study primarily focuses on improving the trajectory planning paradigm and the quality of its planned trajectories given well-perceived environmental information, rather than addressing the perception challenge itself. In real-world scenarios, high-fidelity perception systems in autonomous vehicles already provide such environmental information, ensuring that our approach remains applicable in practical deployments.