acceptodds
Under review as a conference paper at ICLR 2027

Sometimes You Gotta Run Before You Can Walk: Run-then-Walk Scheduling Strategy for VLM Autonomous Driving

Abstract

Recent VLM-based autonomous driving planners adopt GRPO-style reinforcement learning to optimize driving performance. However, existing GRPO recipes either optimize driving efficiency, risking progress-seeking but unsafe behavior, or enforce early safety constraints, leading to overly conservative behavior; both require lengthy training. To solve these problems, we first reveal two distinct RL regimes: a progress regime (Run-GRPO) that aggressively explores high progress, and a safety regime (Walk-GRPO) that restores safety under stable progress. Based on this finding, we propose , a simple yet effective two-stage reward scheduling strategy for GRPO, achieving both better performance and faster convergence. Unlike one-stage RL, which may focus on progress, safety, or a mixture of both within a single training phase, this schedule explicitly separates progress discovery from safety repair. In the phase, we focus on progress, allowing the policy to escape the conservative bias and discover high-progress modes. In the subsequent phase, we introduce endpoint and safety strategy to repair unsafe behaviors from the Run phase. This reversed schedule overcomes the conservatism of Walk-first methods and the unsafe progress-seeking of joint optimization. We validate it with various VLM-based planners on multiple benchmarks: NAVSIMv1, NAVSIMv2, Navhard, and nuScenes. Extensive experiments demonstrate improved driving performance while requiring 40–50 fewer RL training epochs than the baselines.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Autonomous Driving, VLA, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_teqaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Run-then-Walk, a two-stage GRPO training strategy for VLM-based autonomous driving that first encourages driving progress under relaxed constraints and subsequently introduces endpoint and safety rewards to refine the policy. The approach aims to overcome overly conservative behavior from safety-first training while avoiding unsafe aggressive driving. Experiments on NAVSIMv1, NAVSIMv2, Navhard, and nuScenes using autoregressive and diffusion-based planners report improved driving performance and 40–50% fewer RL training epochs compared to existing approaches.

Critical strengths
  1. The paper addresses an important trade-off between driving progress and safety, proposing a simple training strategy that can be integrated into existing VLM planners without architectural modifications.
  2. Experiments across multiple benchmarks and two different planner architectures demonstrate consistent performance improvements with fewer RL epochs. The recipe ablations, training curves, and multi-seed evaluations provide useful supporting evidence.
Critical weaknesses
  1. The methodological novelty is somewhat limited, as the approach primarily changes the ordering of existing reward objectives. The theoretical analysis is highly idealized and does not establish that this particular ordering is optimal or guarantees improved GRPO training.
  2. The claimed training efficiency is not fully validated through controlled comparisons. The proposed method uses 5–6 epochs versus 10 for the original baselines, but equal-budget comparisons, wall-clock measurements, and sensitivity to the stage-switching point are insufficiently explored.
  3. The improvements are relatively modest on some benchmarks, and the safety trade-off is not uniformly positive. For instance, AutoDrive-P3 achieves higher PDMS on NAVSIMv1 but slightly lower NC, DAC, and TTC than the original baseline. Additionally, changes in learning rate and reference policy between stages are not independently ablated, making it difficult to isolate the effect of reward scheduling.
Questions for authors
  1. Can the authors provide equal-compute comparisons against the original training strategies and report actual GPU-hours or wall-clock training costs?
  2. How sensitive is performance to the Run/Walk epoch allocation and switching criterion? Would alternative schedules or a gradual reward transition achieve similar results?
  3. Can the authors independently evaluate the effects of reward ordering, reference-policy resetting, and stage-specific learning rates, particularly for ReCogDrive?
  4. How do the authors justify the claimed safety improvements when some safety metrics decrease relative to the original baseline? Are these differences acceptable in safety-critical driving scenarios?
10 comments

Thank you for your valuable and insightful comments. We deeply appreciate your input and have made every effort to address your concerns, as detailed below.

Response to W1: Novelty and Theory (1/2)

We apologize for the misunderstanding of our contributions arising from the wording of our original manuscript. Run-then-Walk is not a simple reordering of existing rewards: it is derived from an RL training phenomenon that previous VLM-based driving work does not identify, and it is selected by systematic recipe comparisons across planners and benchmarks.

(1) Contributions. First, we identify two RL regimes for VLM-based driving from the same SFT checkpoint (Figs. 3 and 5 of the original manuscript): a progress regime (Run-GRPO) that explores high progress at the cost of safety, and a safety regime (Walk-GRPO) that preserves safety but stays conservative. Second, based on this finding, Run-then-Walk first expands the progress distribution with a PDMS-only Run stage and then repairs safety with a gated endpoint-and-safety Walk stage whose KL reference is reset to the Run checkpoint; this differs from the Walk-and-Run recipe of ReCogDrive [1], the Walk-then-Run recipe of AutoDrive-P3 [2] and generic easy-to-hard curricula [3], which all start from constrained or easy objectives (Figure 1). Third, Run-then-Walk is the best recipe on both planners (Tables 1 and 2), improves performance on four benchmarks and saves 40–50% of RL epochs, and applies to both autoregressive and diffusion VLM planners without changing the planner architecture or the inference cost.

(2) Theory. The role of the theory is to explain why the Run-then-Walk direction works: as stated in Section 4.3, it "serves as a reward-direction explanation rather than a convergence guarantee". Its conditions can be checked against our experiments, and the theory, Figs. 3 and 5, and Table 5 of the original manuscript are consistent with each other (Table 3). The bound J ≤ min{E[S], E[Q]} explains why neither progress alone nor safety alone suffices; the positive reward gap of the Walk gate explains why Walk restores safety from the Run checkpoint; and the progress-retention bound explains why Run must come first: Walk can only retain the high-progress mass that already exists in its reference policy, and Run increases this mass from 46.2% to 85.1% of scenes. We have made this connection explicit in Appendix B of the updated manuscript.

Figure 1. Compared RL training recipes

                  start of RL ──────────────────────────────────────▶ end of RL

                  ┌───────────────────────────────────────────────────────────┐
Walk-then-Run     │   strict safety ───── gradually relaxed ─────▶ progress   │  10 epochs
(AutoDrive-P3)    └───────────────────────────────────────────────────────────┘
                  ┌───────────────────────────────────────────────────────────┐
Walk-and-Run      │   progress and safety objectives combined in one stage    │  10 epochs
(ReCogDrive)      └───────────────────────────────────────────────────────────┘
                  ┌───────────────────────────────────────────────────────────┐
Linear annealing  │ R_plan^run ───── weight moves linearly ─────▶ R_plan^walk │  5 / 6 epochs
(added baseline)  └───────────────────────────────────────────────────────────┘
                  ┌──────────────────────┐   ┌────────────────────────────────┐
Run-then-Walk     │ Run (progress)       │──▶│ Walk (safety repair)           │  5 / 6 epochs
(ours)            │ R_plan = PDMS        │   │ R_plan = I_safe·(NC+DAC+R_end) │
                  │ KL ref.: SFT         │   │ KL ref.: Run checkpoint        │
                  └──────────────────────┘ ▲ └────────────────────────────────┘
                                           │
                  switch when the training reward (PDMS) and its EP component on navtrain plateau

Epochs: AutoDrive-P3 / ReCogDrive.

[1] ReCogDrive: A Reinforced Cognitive Framework for End-to-End Autonomous Driving. (26' ICLR)

[2] AutoDrive-P3: Unified Chain of Perception-Prediction-Planning Thought via Reinforcement Fine-Tuning. (26' ICLR)

[3] Curriculum Learning. (09' ICML)

Response to W1: Novelty and Theory (2/2)

Table 1. RL recipes on NAVSIMv1 (AutoDrive-P3, same SFT checkpoint)

RecipeRL epochsNC↑DAC↑EP↑TTC↑Comf↑PDMS↑
Only Run594.895.089.084.7100.085.8
Only Walk599.197.781.597.3100.089.7
Walk-and-Run598.094.285.494.0100.085.9
Linear annealing (Run → Walk)597.196.388.595.3100.090.5
Walk-then-Run (original AutoDrive-P3)599.297.880.597.4100.089.2
Walk-then-Run (original AutoDrive-P3)1098.997.783.796.699.990.2
Run-then-Walk (ours)598.697.388.995.5100.091.8

Table 2. RL recipes on NAVSIMv1 (ReCogDrive, same SFT checkpoint)

RecipeRL epochsNC↑DAC↑EP↑TTC↑Comf↑PDMS↑
Only Run697.997.686.595.0100.090.4
Only Walk699.098.378.897.699.989.0
Linear annealing (Run → Walk)698.597.785.495.0100.090.7
Walk-then-Run697.996.584.594.699.988.3
Walk-and-Run (original ReCogDrive)698.198.185.495.1100.090.2
Walk-and-Run (original ReCogDrive)1098.197.786.594.9100.090.6
Run-then-Walk (ours)698.697.985.995.8100.091.0

Table 3. Theoretical conditions vs. observations (AutoDrive-P3, NAVSIMv1)

Theoretical statementImplicationObservation
J ≤ min{E[S], E[Q]} (Eq. 8)Neither progress nor safety alone sufficesOnly Run: EP 89.0 but PDMS 85.8; Only Walk: NC 99.1 but PDMS 89.7
Safety direction of Walk (Eqs. 18–19)A positive reward gap γ = 1.5·λ_plan − ω raises the probability of passing the gateγ ≥ 0.85; from the Run checkpoint, Walk raises NC / DAC / TTC by +3.8 / +2.2 / +10.7
Progress retention (Eq. 9)Walk retains the high-progress mass a_R of its reference policyScenes with EP > 0.9: SFT 46.2% → Run 85.1%; EP after Walk 88.9 (reference = Run) vs. 81.5 (reference = SFT, Only Walk)

Response to W2 & Q1: Equal-compute Comparison and GPU-hours

(1) Equal-compute comparison: Run-then-Walk clearly outperforms the original recipes. All hyperparameters follow the original open-source configurations of AutoDrive-P3 and ReCogDrive without additional tuning, and the original recipes are re-trained with the same infrastructure (bf16, 8×A100), SFT checkpoints and seeds; we only change the planning reward and its schedule. The training code and configurations are available in our anonymous repository: https://anonymous.4open.science/r/Run-then-Walk-065A/. In all recipes, each RL epoch samples G=8 rollouts per query and evaluates every rollout with the same PDM scorer, so the same number of RL epochs corresponds to the same compute. Following your suggestion, we re-train both original recipes with the same 5/6 RL epochs as Run-then-Walk (Table 4; numbers in parentheses: PDMS change relative to the original recipe with the same compute). With the same 1,000 GPU-hours, Run-then-Walk improves the PDMS of AutoDrive-P3 by +2.6 (91.8 vs. 89.2); with the same 96 GPU-hours, it improves the PDMS of ReCogDrive by +0.8 (91.0 vs. 90.2). Even with the full 10-epoch compute, the original recipes (90.2 and 90.6) remain below Run-then-Walk, while extending Run-then-Walk to 10 epochs brings no further gain.

(2) GPU-hours and wall-clock cost. One RL epoch takes about 25 h on 8×A100 for AutoDrive-P3 (803 steps), which decodes its perception, prediction and planning tokens autoregressively, and about 2 h for ReCogDrive (1330 steps), whose diffusion head uses 5 DDIM denoising steps as in the original implementation. Run-then-Walk therefore reduces the RL training time from about 250 h (2,000 GPU-hours) to 125 h (1,000 GPU-hours) for AutoDrive-P3 and from about 20 h (160 GPU-hours) to 12 h (96 GPU-hours) for ReCogDrive: the 40–50% reduction holds for the measured GPU-hours, not only for the epoch count.

(3) Switching point. The sensitivity to the switching point is reported in (Sec. 6, Table 7 of updated manuscript).

Table 4. Equal-budget comparison and training cost on NAVSIMv1 (8×A100)

PlannerRecipeRL epochsRL stepsWall-clock (h)GPU-hoursPDMS↑
AutoDrive-P3Walk-then-Run (original)54,0151251,00089.2
AutoDrive-P3Run-then-Walk (3 + 2, ours)54,0151251,00091.8 (+2.6)
AutoDrive-P3Walk-then-Run (original)108,0302502,00090.2
AutoDrive-P3Run-then-Walk (3 + 7)108,0302502,00091.8 (+1.6)
ReCogDriveWalk-and-Run (original)67,980129690.2
ReCogDriveRun-then-Walk (5 + 1, ours)67,980129691.0 (+0.8)
ReCogDriveWalk-and-Run (original)1013,3002016090.6
ReCogDriveRun-then-Walk (5 + 5)1013,3002016091.0 (+0.4)
Reviewer_589TOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies the order of reward objectives during GRPO post-training of VLM driving planners. Run-then-Walk first optimizes PDMS to discover higher-progress trajectories, then switches to an endpoint-consistency reward combined with collision and drivable-area terms. The second phase also resets the KL reference to the Run checkpoint, aiming to preserve its useful progress behavior while repairing safety failures. The recipe is applied to autoregressive AutoDrive-P3 and diffusion-based ReCogDrive without changing their inference architectures.

The reported schedule uses three Run plus two Walk epochs for AutoDrive-P3, and five Run plus one Walk epoch for ReCogDrive. NAVSIMv1 PDMS improves from 90.2 to 91.8 and from 90.6 to 91.0, respectively, compared with ten-epoch original recipes. Additional results cover NAVSIMv2, Navhard, and nuScenes. The supplement includes three-training-seed and three-test-seed results on NAVSIMv1 and an explicitly qualified theoretical analysis of safety reweighting and progress retention.

Critical strengths
  1. The intervention is simple and practical. Changing training objectives and their ordering introduces no additional inference module, making the recipe potentially easy to apply to existing planners.

  2. The paper goes beyond final aggregate scores by examining progress, safety, and trajectory-error dynamics. Figures 3 and 5 show a meaningful transition: the Run stage increases progress while reducing validation safety, followed by recovery during Walk.

  3. Testing both autoregressive and diffusion planners is useful evidence that the recipe is not tied to one output representation. Improvements on multiple evaluation protocols further strengthen the empirical case, even though the NAVSIM variants are related benchmarks.

  4. Table 5 compares several objective orderings, and Table 6 examines safety-reward composition. Appendix D's seed studies support the stability of the NAVSIMv1 paired gains; the paper should receive credit for these existing uncertainty estimates.

  5. Section 4.3 and Appendix B clearly qualify the theory. They acknowledge that useful safe-progress support, achieved KL, and finite group coverage matter, and that endpoint proximity is not a safety certificate.

Critical weaknesses
  1. Higher aggregate scores do not imply uniform safety improvement. For AutoDrive-P3 on NAVSIMv1, PDMS rises from 90.2 to 91.8 while NC falls from 98.9 to 98.6, DAC from 97.7 to 97.3, and TTC from 96.6 to 95.5 (Table 1). Appendix D reproduces these differences across seeds. In contrast, ReCogDrive improves its safety submetrics while slightly reducing progress. The interpretation should distinguish these two outcomes and explicitly characterize the safety/progress trade-off.

  2. Reward ordering is not fully isolated from other training changes. The method combines objective order, a new endpoint reward, a safety gate, and a reset of the KL reference. ReCogDrive additionally changes learning rate from 10−410^{-4} in Run to 10−610^{-6} in Walk (Table 7). Table 5 is helpful, but does not fully specify matched optimization budgets, reference resets, or tuning effort for every recipe. These factors could account for part of the advantage attributed to Run-first scheduling.

  3. The 40–50% reduction is established in epochs, not complete compute cost. The paper compares five/six epochs with ten, but does not report end-to-end GPU-hours, generated samples, reward-evaluation cost, or checkpoint/schedule search overhead. Since the claimed contribution includes faster convergence, equal-compute and equal-sample comparisons are important. Appendix C's decision to switch once progress has plateaued also needs an operational criterion that can be applied without retrospective judgment.

  4. The safety gate and endpoint signal require more scrutiny. The gate permits NC greater than zero, including NC = 0.5, and does not directly enforce TTC or all EPDMS compliance factors (Section 4.1; Appendix B). With Δ=10\Delta=10, the endpoint reward is constant below an error of 2Δ2\Delta, so the proportion of informative versus saturated endpoint rewards should be reported. Aggregate improvements do not establish that the resulting policy consistently repairs every relevant failure mode.

  5. The theoretical conditions are not connected quantitatively to the runs. The progress-retention bound requires achieved full-output KL, and the safety-direction argument requires a reward gap relative to auxiliary-reward variation. These quantities are not measured. Table 7 also omits the Run/Walk KL coefficients and planning-reward weight. The complete-sample likelihood ratio and KL implementation for a diffusion planner need specification. The statement that removing KL causes collapse would be stronger with a quantitative ablation.

  6. Evidence for broader driving robustness remains limited. The principal experiments rely on the NAVSIM family and nuScenes planning evaluation. They do not establish that the same ordering improves long interactive rollouts with reactive traffic. Seed analyses are supplied for NAVSIMv1, but uncertainty for the harder benchmarks and the different objective-ordering ablations remains unclear.

Questions for authors
  1. Add a matched-budget factorial comparison of objective ordering and KL-reference reset. Keep endpoint rewards, learning-rate schedules, sample counts, and tuning budgets identical where applicable. Include the original recipe at five/six epochs and Run-then-Walk at ten epochs.

  2. Report GPU-hours and sampled trajectories required to reach a fixed validation score. Define the progress-plateau switching rule, disclose the data used for selecting it, and test earlier/later switches. This would turn an empirical schedule into a reproducible procedure.

  3. Plot safety/progress frontiers and report per-scenario regressions as well as gains. For AutoDrive-P3, quantify the extra safety violations associated with the higher progress, including uncertainty on harder benchmarks and interactive closed-loop tests.

  4. Ablate the endpoint term separately from the gate and test sensitivity to Δ\Delta and η\eta. State their physical units and report the fraction of samples in each endpoint-reward bin. Compare NC > 0 with a strict NC = 1 gate and examine TTC/DDC/TLC failures.

  5. Provide βrun\beta_{\mathrm{run}}, βwalk\beta_{\mathrm{walk}}, λplan\lambda_{\mathrm{plan}}, auxiliary-reward scales, and the exact likelihood/KL estimators for both planner families. Measure achieved KL and useful safe-progress coverage during Walk, including groups with no passing sample.

  6. Expand the schedule-ordering controls and uncertainty analysis to ReCogDrive and the hard benchmarks. Qualify the main claim as a training recipe that improves the reported aggregate trade-offs with fewer epochs; reserve stronger safety and compute-efficiency claims for direct measurements.

14 comments

Thank you for your valuable and insightful comments. We deeply appreciate your input and have made every effort to address your concerns, as detailed below.

Response to W1 & Q3: Safety–Progress Trade-off, Frontier, and Evaluation Protocol (1/2)

We apologize for the misunderstanding arising from the wording of our original manuscript. As you point out, the two planners show two different outcomes: compared with their original recipes, Run-then-Walk makes the over-aggressive ReCogDrive safer and the over-conservative AutoDrive-P3 much more efficient, and it improves PDMS and EPDMS for both.

(1) Two outcomes from two opposite biases. As illustrated in Fig. 2 of the original manuscript, AutoDrive-P3 [1] follows a Walk-then-Run recipe that is safe but overly conservative (EP 83.7, below the human reference of 87.5); Run-then-Walk trades a small amount of safety (NC −0.3, DAC −0.4, TTC −1.1) for a large progress gain (EP +5.2, PDMS 90.2 → 91.8). ReCogDrive [2] follows a Walk-and-Run recipe biased toward progress; Run-then-Walk trades a small amount of progress (EP −0.6) for consistent safety gains (NC +0.5, DAC +0.2, TTC +0.9; PDMS 90.6 → 91.0; Table 1). The same pattern holds on NAVSIMv2 (Table 2) and Navhard (EPDMS 29.3 → 31.3 for AutoDrive-P3 and 26.5 → 28.2 for ReCogDrive). We have revised Section 1 and Section 5 to state these two outcomes separately.

(2) Safety/progress frontier. Following your suggestion, Table 3 lists the safety/progress operating points of AutoDrive-P3 evaluated in the manuscript, sorted by EP. In both the EP–NC and EP–TTC planes, the original recipe and the two Run-then-Walk variants (Walk reward DAC+NC and DAC+NC+TTC) lie on the non-dominated frontier, and Run-then-Walk extends this frontier toward higher EP. The trade-off can also be chosen explicitly: with TTC added to the Walk reward, NC, DAC and TTC stay within 0.6 of the original recipe, while EP still improves by 3.2 and PDMS by 1.0.

(3) Evaluation protocol. Our evaluation strictly follows the official NAVSIM protocol: NAVSIMv1, NAVSIMv2 and Navhard evaluate a method by its metrics averaged over all test scenes (more than 12k on navtest, so reporting every scene individually is not feasible), and AutoDrive-P3, ReCogDrive and the other published methods are reported in the same way. The extra safety violations of AutoDrive-P3 are therefore quantified directly by the official NC, DAC and TTC differences in Table 1 (−0.3, −0.4 and −1.1), against an EP gain of +5.2.

(4) Harder benchmarks and closed-loop tests. NAVSIMv2 and Navhard are the harder benchmarks and closed-loop tests: NAVSIMv2 adds DDC, TLC, LK, HC and EC to the score, and Navhard is a two-stage pseudo closed-loop evaluation in which the planner re-plans from synthetic start states around its own 4 s endpoint. In the second stage of Navhard, the NC and DAC of AutoDrive-P3 improve (+0.4 and +3.8; Table 3 of updated manuscript) and its EPDMS rises from 29.3 to 31.3. The training-seed results on NAVSIMv2 (Appendix E, Table 22 of updated manuscript) show that these gains are robust.

[1] AutoDrive-P3: Unified Chain of Perception-Prediction-Planning Thought via Reinforcement Fine-Tuning. (26' ICLR)

[2] ReCogDrive: A Reinforced Cognitive Framework for End-to-End Autonomous Driving. (26' ICLR)

Response to W1 & Q3: Safety–Progress Trade-off, Frontier, and Evaluation Protocol (2/2)

Table 1. Original recipes vs. Run-then-Walk on NAVSIMv1

MethodNC↑DAC↑EP↑TTC↑Comf↑PDMS↑
AutoDrive-P3 (Walk-then-Run)98.997.783.796.699.990.2
AutoDrive-P3 (Run-then-Walk, ours)98.6 (−0.3)97.3 (−0.4)88.9 (+5.2)95.5 (−1.1)100.0 (+0.1)91.8 (+1.6)
ReCogDrive (Walk-and-Run)98.197.786.594.9100.090.6
ReCogDrive (Run-then-Walk, ours)98.6 (+0.5)97.9 (+0.2)85.9 (−0.6)95.8 (+0.9)100.0 (0.0)91.0 (+0.4)

Table 2. Original recipes vs. Run-then-Walk on NAVSIMv2

MethodNC↑DAC↑DDC↑TLC↑EP↑TTC↑LK↑HC↑EC↑EPDMS↑
AutoDrive-P3 (Walk-then-Run)98.997.698.999.886.898.595.498.380.688.7
AutoDrive-P3 (Run-then-Walk, ours)98.6 (−0.3)97.3 (−0.3)98.7 (−0.2)99.7 (−0.1)92.8 (+6.0)98.1 (−0.4)96.1 (+0.7)97.5 (−0.8)78.3 (−2.3)89.6 (+0.9)
ReCogDrive (Walk-and-Run)98.297.798.4100.089.997.490.897.726.882.7
ReCogDrive (Run-then-Walk, ours)98.5 (+0.3)97.9 (+0.2)98.8 (+0.4)100.0 (0.0)89.4 (−0.5)97.8 (+0.4)92.0 (+1.2)98.2 (+0.5)29.8 (+3.0)83.7 (+1.0)

Table 3. Safety/progress operating points of AutoDrive-P3 on NAVSIMv1 (sorted by EP)

PolicyRL epochsEP↑NC↑DAC↑TTC↑PDMS↑
Only Walk581.599.197.797.389.7
SFT081.798.695.795.387.8
Walk-then-Run (original)1083.798.997.796.690.2
Walk-and-Run585.498.094.294.085.9
Run-then-Walk, Walk reward DAC+NC+TTC586.998.797.496.091.2
Run-then-Walk, Walk reward DAC588.098.597.194.991.5
Run checkpoint388.894.895.184.885.6
Run-then-Walk, Walk reward DAC+NC (ours)588.998.697.395.591.8

Response to W2 & Q1: Isolating Reward Ordering from Other Training Changes (1/2)

(1) No extra tuning, and the learning-rate typo. All hyperparameters, including learning rate, batch size, group size, optimizer and KL settings, follow the original open-source configurations of AutoDrive-P3 and ReCogDrive without additional tuning, with Δ=10 m and η=0.2 for both planners; we only change the planning reward and its training order. The gains are stable across three training seeds and three test seeds (Tables 8 and 9 of the original manuscript) and across training seeds on NAVSIMv2 (Appendix E, Table 22 of updated manuscript). The learning rate of 1×10⁻⁶ in the ReCogDrive Walk stage (Table 7 of the original manuscript) is a typo: ReCogDrive uses 1×10⁻⁴ in both Run and Walk; we have corrected it in the manuscript. Running the ReCogDrive Walk stage with 1×10⁻⁶ gives a similar result (Table 4): PDMS 91.0 vs. 91.0, NAVSIMv2 EPDMS 83.6 vs. 83.7, and Navhard EPDMS 28.2 vs. 28.2.

(2) Equal-epoch budget. With the same 5/6-epoch budget, the original recipes are clearly below Run-then-Walk (Table 8, including the training cost; Sec. 6, Table 6 of updated manuscript).

(3) RL recipes with all other settings fixed. Table 6 compares the RL recipes of AutoDrive-P3 from the same SFT checkpoint (Table 5 of the original manuscript), including Walk-and-Run, which optimizes the two objectives jointly as a fixed reward mixture. We add linear annealing and repeat the comparison on ReCogDrive (Table 7). Run-then-Walk is the best recipe on both planners (Figure 1 illustrates the compared recipes).

(4) KL regularization and reference reset. Table 5 keeps the SFT model as the reference of the Walk stage, with the same rewards, learning rate, sampled trajectories and seeds: the policy stays constrained by the SFT model and reaches only EP 86.9 and PDMS 90.9 (vs. 88.9 and 91.8). Reversing the order is clearly worse for both planners (Walk-then-Run in Tables 6 and 7), and removing the KL term makes the training of AutoDrive-P3 unstable, with PDMS 86.4, below the SFT model (Appendix E, Table 20 of updated manuscript). These results show that the gain comes from the two RL regimes we identify, the progress regime of Run-GRPO and the safety regime of Walk-GRPO (Figs. 3 and 5 of the original manuscript), used in the Run-then-Walk order.

Reviewer_u9AJOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a two-stage GRPO reward schedule for VLM driving planners. Run optimizes PDMS directly; Walk switches to a safety-gated endpoint reward and uses the Run checkpoint as the KL reference. Experiments with AutoDrive-P3 and ReCogDrive report modest benchmark improvements using five to six RL epochs instead of ten, alongside training-dynamics analyses and an idealized KL-regularized argument.

Critical strengths
  1. The schedule is simple and is demonstrated with both autoregressive and diffusion planners. If the training-cost advantage holds under matched conditions, this would be a useful practical improvement that does not require changing the planner architecture.

  2. The training-dynamics plots make the progress/safety trade-off concrete, and the appendix includes multiple training and test seeds. This is useful evidence for studying reward scheduling rather than relying only on a final aggregate score.

Critical weaknesses
  1. The safety interpretation is stronger than the baseline comparison supports. For AutoDrive-P3 on NAVSIMv1, NC decreases from 98.9 to 98.6, DAC from 97.7 to 97.3, and TTC from 96.6 to 95.5, while EP increases from 83.7 to 88.9. Several safety-related metrics also decrease on NAVSIMv2. Walk may recover safety relative to the Run checkpoint, but this does not establish safer behavior than the original recipe. The paper needs to distinguish these comparisons and characterize the trade-off behind the aggregate gain.

  2. It is not clear that the schedule itself, rather than budget or tuning choices, accounts for the improvement. Reward curricula are established, the aggregate gains are small, and simple matched-budget alternatives such as a fixed reward mixture or linear annealing are absent. The switching criterion and the different epoch splits require a clearly independent selection protocol. Comparing five to six proposed epochs with ten baseline epochs alone does not establish a training-efficiency advantage. These limitations leave me marginally below acceptance.

Questions for authors
  1. Which data split and criterion determined the Run-to-Walk switch and the planner-specific epoch splits? Were all choices fixed before evaluating navtest? Please clarify the protocol rather than assuming test-set tuning occurred.

  2. Were the baselines re-trained under the same infrastructure and seeds? What are their scores at the same five-to-six-epoch budget, and how do wall-clock/GPU-hour costs compare? A fixed mixture or linear transition between the two rewards would help isolate the value of the discrete order.

  3. How much safety is recovered from the Run checkpoint, and how does the final policy compare with the original recipe per scenario? Since PDMS already includes safety factors, why does directly optimizing it reduce validation PDMS during Run? Please also clarify what randomness varies across seeds.

11 comments

Thank you for your valuable and insightful comments. We deeply appreciate your input and have made every effort to address your concerns, as detailed below.

Response to W1: Two Outcomes on Two Planners

We apologize for the misunderstanding arising from the wording of our original manuscript. Compared with the original recipes, Run-then-Walk improves the overall performance (PDMS and EPDMS) of both planners in two different ways: it makes the over-aggressive ReCogDrive safer, and it makes the over-conservative AutoDrive-P3 much more efficient while keeping it relatively safe.

(1) The two original recipes are biased in opposite directions. As illustrated in Fig. 2 of the original manuscript, AutoDrive-P3 [1] follows a Walk-then-Run recipe: it starts from strict safety constraints and only gradually relaxes them, so the policy explores only around the conservative SFT solution. It is safe (NC 98.9, TTC 96.6) but inefficient (EP 83.7, below the human reference of 87.5; Figs. 6, 9 and 10 of the original manuscript). ReCogDrive [2] follows a Walk-and-Run recipe: it combines progress and safety in a single stage, and its training remains biased toward progress. It is efficient (EP 86.5) but less safe (NC 98.1, TTC 94.9; Figs. 11 and 12 of the original manuscript).

(2) AutoDrive-P3: a large progress gain at a small safety cost. With better exploration in the Run stage, Run-then-Walk moves the conservative policy toward a better balance: at a small safety cost (NC −0.3, DAC −0.4, TTC −1.1), EP rises by +5.2 and PDMS from 90.2 to 91.8 on NAVSIMv1 (Table 1), and EP rises by +6.0 and EPDMS from 88.7 to 89.6 on NAVSIMv2 (Table 2). Its NC of 98.6 remains higher than that of most compared methods, such as ReCogDrive (98.1), DiffusionDrive (98.2) and WoTE (98.5).

(3) ReCogDrive: consistent safety gains at a small progress cost. With the safety repair in the Walk stage, Run-then-Walk makes the progress-biased policy safer: NC, DAC and TTC improve by +0.5, +0.2 and +0.9 on NAVSIMv1 at an EP cost of 0.6, and PDMS rises from 90.6 to 91.0 (Table 1); on NAVSIMv2, EPDMS rises from 82.7 to 83.7 (Table 2).

(4) Revision. We have revised the sentence in Section 1, "The final policies retain high EP and restore collision, drivable-area, and time-to-collision safety metrics", to "Relative to the original recipes, the final policies achieve higher safety for the progress-biased ReCogDrive and higher progress for the conservative AutoDrive-P3", and revised the discussion in Section 5 accordingly. The recovery from the Run checkpoint is reported in (Sec. 6, Table 9 of updated manuscript).

Table 1. Original recipes vs. Run-then-Walk on NAVSIMv1

MethodNC↑DAC↑EP↑TTC↑Comf↑PDMS↑
AutoDrive-P3 (Walk-then-Run)98.997.783.796.699.990.2
AutoDrive-P3 (Run-then-Walk, ours)98.6 (−0.3)97.3 (−0.4)88.9 (+5.2)95.5 (−1.1)100.0 (+0.1)91.8 (+1.6)
ReCogDrive (Walk-and-Run)98.197.786.594.9100.090.6
ReCogDrive (Run-then-Walk, ours)98.6 (+0.5)97.9 (+0.2)85.9 (−0.6)95.8 (+0.9)100.0 (0.0)91.0 (+0.4)

Table 2. Original recipes vs. Run-then-Walk on NAVSIMv2

MethodNC↑DAC↑DDC↑TLC↑EP↑TTC↑LK↑HC↑EC↑EPDMS↑
AutoDrive-P3 (Walk-then-Run)98.997.698.999.886.898.595.498.380.688.7
AutoDrive-P3 (Run-then-Walk, ours)98.6 (−0.3)97.3 (−0.3)98.7 (−0.2)99.7 (−0.1)92.8 (+6.0)98.1 (−0.4)96.1 (+0.7)97.5 (−0.8)78.3 (−2.3)89.6 (+0.9)
ReCogDrive (Walk-and-Run)98.297.798.4100.089.997.490.897.726.882.7
ReCogDrive (Run-then-Walk, ours)98.5 (+0.3)97.9 (+0.2)98.8 (+0.4)100.0 (0.0)89.4 (−0.5)97.8 (+0.4)92.0 (+1.2)98.2 (+0.5)29.8 (+3.0)83.7 (+1.0)

[1] AutoDrive-P3: Unified Chain of Perception-Prediction-Planning Thought via Reinforcement Fine-Tuning. (26' ICLR)

[2] ReCogDrive: A Reinforced Cognitive Framework for End-to-End Autonomous Driving. (26' ICLR)

Response to W2 & Q2: Schedule vs. Budget and Tuning (1/2)

(1) Fixed reward mixture and linear annealing. Table 5 of the original manuscript compares different objective orderings from the same AutoDrive-P3 SFT checkpoint, including Walk-and-Run, which optimizes the Run and Walk objectives jointly as a fixed reward mixture. Following your suggestion, we add linear annealing, where the planning reward moves linearly from the Run reward to the Walk reward under the same budget, R_plan(t) = (1 − t/T)·R_plan^run + (t/T)·R_plan^walk, and repeat the full recipe comparison on ReCogDrive (Figure 1, Tables 3 and 4). Run-then-Walk remains the best recipe for both planners. The gain therefore comes from the two RL regimes we identify, a progress regime under the Run reward and a safety regime under the Walk reward (Figs. 3 and 5 of the original manuscript), used in the Run-then-Walk order, rather than from combining the two objectives.

(2) No extra tuning, and the same infrastructure and seeds for the baselines. All hyperparameters, including learning rate, batch size, group size, optimizer and KL settings, follow the original open-source configurations of AutoDrive-P3 and ReCogDrive; we only change the planning reward and its schedule, with the endpoint parameters fixed to Δ=10 m and η=0.2 for both planners. The original recipes are re-trained with the same infrastructure (bf16, 8×A100), SFT checkpoints and three training seeds as ours. Over the three training seeds, PDMS is 91.78 ± 0.03 vs. 90.18 ± 0.06 for AutoDrive-P3 and 91.02 ± 0.10 vs. 90.58 ± 0.12 for ReCogDrive, far larger gains than the seed variation (Tables 8 and 9 of the original manuscript). The training code and configurations are available in our anonymous repository: https://anonymous.4open.science/r/Run-then-Walk-065A/.

(3) Same 5/6-epoch budget. We re-train both original recipes with the same 5/6 RL epochs as Run-then-Walk (Tables 3–5). Under the same budget, Run-then-Walk improves PDMS by +2.6 for AutoDrive-P3 (91.8 vs. 89.2) and by +0.8 for ReCogDrive (91.0 vs. 90.2). Even with the full 10 epochs, the original recipes (90.2 and 90.6) remain below Run-then-Walk with 5/6 epochs, while extending Run-then-Walk to 10 epochs brings no further gain. The improvement therefore comes from the Run-then-Walk order of the two regimes rather than from the training budget.

(4) Wall-clock time and GPU-hours. In all recipes, each RL epoch samples G=8 rollouts per query and scores them with the same PDM scorer, so the per-epoch cost is the same for all recipes of a planner (Table 5). One RL epoch takes about 25 h on 8×A100 for AutoDrive-P3 (803 steps), which decodes its perception, prediction and planning tokens autoregressively, and about 2 h for ReCogDrive (1330 steps), whose diffusion head uses 5 DDIM denoising steps as in the original implementation. Run-then-Walk reduces the RL training time from about 250 h (2,000 GPU-hours) to 125 h (1,000 GPU-hours) for AutoDrive-P3 and from about 20 h (160 GPU-hours) to 12 h (96 GPU-hours) for ReCogDrive, so the 40–50% reduction holds for the measured training cost.

(5) Comparison with published methods. The AutoDrive-P3 variant achieves the best scores among the compared methods on NAVSIMv1, NAVSIMv2 and Navhard (Table 6).

The switching protocol is described in (Sec. 4.2, Appendix C and Algorithm 1 of updated manuscript).

Figure 1. Compared RL training recipes

                  start of RL ──────────────────────────────────────▶ end of RL

                  ┌───────────────────────────────────────────────────────────┐
Walk-then-Run     │   strict safety ───── gradually relaxed ─────▶ progress   │  10 epochs
(AutoDrive-P3)    └───────────────────────────────────────────────────────────┘
                  ┌───────────────────────────────────────────────────────────┐
Walk-and-Run      │   progress and safety objectives combined in one stage    │  10 epochs
(ReCogDrive)      └───────────────────────────────────────────────────────────┘
                  ┌───────────────────────────────────────────────────────────┐
Linear annealing  │ R_plan^run ───── weight moves linearly ─────▶ R_plan^walk │  5 / 6 epochs
(added baseline)  └───────────────────────────────────────────────────────────┘
                  ┌──────────────────────┐   ┌────────────────────────────────┐
Run-then-Walk     │ Run (progress)       │──▶│ Walk (safety repair)           │  5 / 6 epochs
(ours)            │ R_plan = PDMS        │   │ R_plan = I_safe·(NC+DAC+R_end) │
                  │ KL ref.: SFT         │   │ KL ref.: Run checkpoint        │
                  └──────────────────────┘ ▲ └────────────────────────────────┘
                                           │
                  switch when the training reward (PDMS) and its EP component on navtrain plateau

Epochs: AutoDrive-P3 / ReCogDrive.

Response to W2 & Q2: Schedule vs. Budget and Tuning (2/2)

Table 3. RL recipes on NAVSIMv1 (AutoDrive-P3, same SFT checkpoint)

RecipeRL epochsNC↑DAC↑EP↑TTC↑Comf↑PDMS↑
Only Run594.895.089.084.7100.085.8
Only Walk599.197.781.597.3100.089.7
Walk-and-Run598.094.285.494.0100.085.9
Linear annealing (Run → Walk)597.196.388.595.3100.090.5
Walk-then-Run (original AutoDrive-P3)599.297.880.597.4100.089.2
Walk-then-Run (original AutoDrive-P3)1098.997.783.796.699.990.2
Run-then-Walk (ours)598.697.388.995.5100.091.8

Table 4. RL recipes on NAVSIMv1 (ReCogDrive, same SFT checkpoint)

RecipeRL epochsNC↑DAC↑EP↑TTC↑Comf↑PDMS↑
Only Run697.997.686.595.0100.090.4
Only Walk699.098.378.897.699.989.0
Linear annealing (Run → Walk)698.597.785.495.0100.090.7
Walk-then-Run697.996.584.594.699.988.3
Walk-and-Run (original ReCogDrive)698.198.185.495.1100.090.2
Walk-and-Run (original ReCogDrive)1098.197.786.594.9100.090.6
Run-then-Walk (ours)698.697.985.995.8100.091.0

Table 5. Equal-budget comparison and training cost on NAVSIMv1 (8×A100)

PlannerRecipeRL epochsRL stepsWall-clock (h)GPU-hoursPDMS↑
AutoDrive-P3Walk-then-Run (original)54,0151251,00089.2
AutoDrive-P3Run-then-Walk (3 + 2, ours)54,0151251,00091.8 (+2.6)
AutoDrive-P3Walk-then-Run (original)108,0302502,00090.2
AutoDrive-P3Run-then-Walk (3 + 7)108,0302502,00091.8 (+1.6)
ReCogDriveWalk-and-Run (original)67,980129690.2
ReCogDriveRun-then-Walk (5 + 1, ours)67,980129691.0 (+0.8)
ReCogDriveWalk-and-Run (original)1013,3002016090.6
ReCogDriveRun-then-Walk (5 + 5)1013,3002016091.0 (+0.4)

Table 6. Comparison with published methods

MethodNAVSIMv1 PDMS↑NAVSIMv2 EPDMS↑Navhard EPDMS↑
DiffusionDrive (25' CVPR)88.188.228.9
WoTE (25' ICCV)88.387.7–
DriveVLA-W0 (25' arXiv)90.286.1–
DriveFuture (26' arXiv)90.7––
DriveWorld-VLA (26' arXiv)91.386.8–
Latent-WAM (26' arXiv)–89.3–
ReCogDrive (26' ICLR)90.682.726.5
ReCogDrive (Run-then-Walk, ours)91.083.728.2
AutoDrive-P3 (26' ICLR)90.288.729.3
AutoDrive-P3 (Run-then-Walk, ours)91.889.631.3

Comments on the paper

2 comments

General Response (1/2)

Dear AC,

We thank the reviewers for their careful reading and constructive comments. Below we summarize the shared concerns, the new evidence and the revisions; the individual responses give the full details. The reviewers find Run-then-Walk simple and applicable to both autoregressive and diffusion planners without architectural changes (u9AJ, 589T, teqa), the training dynamics a meaningful view of the progress/safety transition (u9AJ, 589T), and the experiments across benchmarks, planners and seeds supportive of consistent gains with fewer RL epochs (589T, teqa). The shared concerns are (A) whether the gain comes from the reward order rather than budget or tuning, (B) how to read the safety–progress trade-off of the two planners, and (C) how the switch is chosen and what the training costs. We address them with equal-budget re-training of both original recipes, linear annealing, the full recipe comparison on ReCogDrive, switching-point, gate, endpoint and KL ablations, and three-seed NAVSIMv2 results. All hyperparameters follow the original open-source configurations of AutoDrive-P3 and ReCogDrive; we change only the planning reward and its order. The code is available at https://anonymous.4open.science/r/Run-then-Walk-065A.

GR1. The reward order, not budget or tuning, drives the gain (u9AJ W2/Q2, 589T W2/W3, teqa W2/Q1). All recipes of a planner start from the same SFT checkpoint and have the same per-epoch cost (Sec. 6, Tables 5 and 6 of updated manuscript).

PlannerRecipeRL epochsGPU-hoursPDMS↑
AutoDrive-P3Walk-then-Run (original)5 / 101,000 / 2,00089.2 / 90.2
AutoDrive-P3Walk-and-Run / Linear annealing51,00085.9 / 90.5
AutoDrive-P3Run-then-Walk (ours)51,00091.8
ReCogDriveWalk-and-Run (original)6 / 1096 / 16090.2 / 90.6
ReCogDriveWalk-then-Run / Linear annealing69688.3 / 90.7
ReCogDriveRun-then-Walk (ours)69691.0

We highlight three points:

Run-then-Walk wins under the same budget. With the same 5/6 RL epochs, it improves PDMS by +2.6 for AutoDrive-P3 and +0.8 for ReCogDrive. The original recipes stay below it even with 10 epochs, so Run-then-Walk reduces the RL cost from 2,000 to 1,000 GPU-hours and from 160 to 96 GPU-hours.

The order matters, not the combination of objectives. A fixed reward mixture (Walk-and-Run), linear annealing and the reversed order (Walk-then-Run) are all below Run-then-Walk on both planners.

The gains exceed seed variation. Over three training seeds, PDMS is 91.78 ± 0.03 vs. 90.18 ± 0.06 for AutoDrive-P3 and 91.02 ± 0.10 vs. 90.58 ± 0.12 for ReCogDrive.

GR2. Two planners, two outcomes, consistent on three benchmarks (u9AJ W1/Q3, 589T W1/Q3/W6, teqa W3/Q4). AutoDrive-P3 (Walk-then-Run) is overly conservative and ReCogDrive (Walk-and-Run) is progress-biased; Run-then-Walk moves both toward a better balance (Tables 1–3 of updated manuscript).

Planner (original / ours)NAVSIMv1 PDMS↑NAVSIMv2 EPDMS↑Navhard EPDMS↑NAVSIMv1 sub-metrics
AutoDrive-P390.2 / 91.888.7 / 89.629.3 / 31.3EP +5.2; NC −0.3, DAC −0.4, TTC −1.1
ReCogDrive90.6 / 91.082.7 / 83.726.5 / 28.2NC +0.5, DAC +0.2, TTC +0.9; EP −0.6

We highlight two points on the safety–progress trade-off:

The safety cost on AutoDrive-P3 is small and controllable. Its NC of 98.6 remains higher than that of most compared methods. Adding TTC to the Walk reward keeps NC, DAC and TTC within 0.6 of the original recipe while EP still improves by 3.2 and PDMS by 1.0, and the original recipe and both Run-then-Walk variants lie on the non-dominated safety–progress frontier (Appendix E, Table 21 of updated manuscript).

Walk repairs the safety lost in Run. From the Run checkpoint, Walk raises NC/DAC/TTC by +3.8/+2.2/+10.7 for AutoDrive-P3 and by +0.7/+0.3/+0.8 for ReCogDrive (Sec. 6, Table 9 of updated manuscript).

General Response (2/2)

Concerns of each reviewer and our responses.

Reviewer u9AJ

  • Gain from budget or tuning rather than the reward order (W2/Q2): under the same 5/6 RL epochs and GPU-hours, Run-then-Walk is +2.6/+0.8 PDMS above the original recipes, and the fixed mixture, linear annealing and the reversed order are all lower (GR1).
  • Safety relative to the original recipe (W1): AutoDrive-P3 trades a small, controllable safety cost for +5.2 EP, and ReCogDrive gains NC, DAC and TTC; the revised Sec. 1 states these two outcomes (GR2).
  • Switch protocol (Q1): the switch uses only the training reward of the Run rollouts on navtrain, never navtest; switching after 1/2/3/4 Run epochs gives 90.2/91.0/91.8/91.8 PDMS (Sec. 6, Table 7 and Appendix C of updated manuscript).
  • Safety recovery and why Run lowers safety (Q3): Walk restores NC/DAC/TTC by +3.8/+2.2/+10.7 from the Run checkpoint (GR2). During Run, EP dominates the group-normalized advantage because it varies continuously while NC and DAC are often identical within a group (Appendix E, Table 14 of updated manuscript).

Reviewer 589T

  • Trade-off and evaluation protocol (W1/Q3): the original recipe and both Run-then-Walk variants lie on the non-dominated safety–progress frontier (GR2). We follow the official NAVSIM protocol, which averages each metric over all test scenes, and Run-then-Walk also improves both planners on NAVSIMv2 and the two-stage Navhard.
  • Reward order vs. other training changes (W2/Q1): the ReCogDrive learning rate is 1×10⁻⁴ in both stages (a typo in Table 7 of the original manuscript, now fixed), and a Walk learning rate of 1×10⁻⁶ gives the same 91.0 PDMS; keeping SFT as the Walk reference reaches only 90.9 vs. 91.8 (Appendix E, Tables 15 and 16 of updated manuscript).
  • Training cost and switching rule (W3/Q2): the RL cost halves (GR1), and the switch uses navtrain training rewards only.
  • Safety gate and endpoint reward (W4/Q4): a strict NC = 1 gate gives 91.3 PDMS, removing the endpoint term gives 90.4, and PDMS rises with the step size Δ to 91.8 at Δ = 10 m (Appendix E, Tables 17–19 of updated manuscript).
  • KL, training settings and theory (W5/Q5): each planner keeps its original KL setting (without KL, 86.4 PDMS); the Walk gate has a reward gap of at least 0.85 > 0, and Run raises the share of scenes with EP > 0.9 from 46.2% to 85.1% (Appendix B, Table 10 and Appendix C, Table 11 of updated manuscript).
  • Uncertainty and ReCogDrive (W6/Q6): over three training seeds on NAVSIMv2, EPDMS is 89.63 ± 0.03 vs. 88.75 ± 0.05 for AutoDrive-P3 and 83.71 ± 0.12 vs. 82.68 ± 0.12 for ReCogDrive (Appendix E, Table 22 of updated manuscript); the full recipe comparison now covers ReCogDrive (GR1).

Reviewer teqa

  • Novelty and theory (W1): we identify two RL regimes of VLM planners from the same SFT checkpoint and order them with a KL-reference reset; the theory explains this direction, and each of its conditions matches a measured quantity (Appendix B, Table 10 of updated manuscript).
  • Equal compute and GPU-hours (W2/Q1): GR1.
  • Magnitude of the improvement and safety trade-off (W3/Q4): the gains exceed seed variation (GR1), and the safety cost of AutoDrive-P3 is small and controllable (GR2).
  • Epoch allocation and gradual transition (Q2): the switch follows the navtrain training reward, and linear annealing stays below Run-then-Walk on both planners (90.5 and 90.7 vs. 91.8 and 91.0).
  • Independent effects of ordering, reference reset and learning rate (Q3): the Walk learning rate does not change the result (91.0), the KL-reference reset adds 0.9 PDMS, and the reversed order on ReCogDrive reaches only 88.3 PDMS.

Revisions in the updated manuscript (highlighted in blue).

  • Sec. 1: the claim now reads "Relative to the original recipes, the final policies achieve higher safety for the progress-biased ReCogDrive and higher progress for the conservative AutoDrive-P3."
  • Sec. 6: recipe comparison with linear annealing on both planners (Table 5), equal-budget comparison and training cost (Table 6), switching point (Table 7), safety reward terms (Table 8) and safety recovery (Table 9).
  • Appendix B: connection between the theory and the experiments (Table 10).
  • Appendix C and Algorithm 1: the navtrain-only switching rule, the KL setting and the training cost; Table 11 completes the training settings and fixes the ReCogDrive learning rate.
  • Appendix E: zero proportions, KL reference, Walk learning rate, NC gate, endpoint term, Δ, KL coefficient, safety/progress operating points and NAVSIMv2 training seeds (Tables 14–22).
  • Abstract and reproducibility statement: anonymous code link.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.