- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 3(+1)
- Presentation
- 3(+1)
- Soundness
- 2
Read review
Paper summary
The paper proposes BEV Action Engram, a conditional lookup memory for flow-based autonomous driving planners. Spatial displacement n-grams and prior-step velocity predictions index into embedding tables whose outputs modulate the action decoder, aiming to preserve planning quality with fewer denoising steps under imitation learning and Flow-GRPO fine-tuning.
Critical strengths
-
The memory architecture is concrete and well-ablated in the imitation learning setting: Table 3 compares against a parameter-matched FFN and systematically ablates branch selection, conditioning signals, and injection depth.
-
At reduced denoising steps (), Table 8 shows that the model reaches 86.64 PDMS compared to 86.19 for the 5-step baseline while cutting planner FLOPs by ~38% and planner latency by ~14% (yielding a ~4.8% end-to-end latency reduction once the visual backbone is included).
Critical weaknesses
-
The rollout-based Flow Matching training objective in Section 3.7 and Algorithm 2 is mathematically underspecified and inconsistent with the appendix. Algorithm 2 queries the velocity field on self-generated rollout states while supervising every step with the constant vector . Under standard conditional Flow Matching, is the conditional vector field along the linear interpolation path , not along an arbitrary off-path rollout trajectory. Moreover, Appendix Eq. 27 (p. 16) states that is sampled continuously during IL, without explaining how continuous time sampling defines sequential integration steps in a -step rollout.
-
Both the efficiency claim and the RL safety gains are confounded by step count and benchmark noise. At any fixed step budget in Table 8, adding Engram substantially increases planner latency ( ms at ; ms at ), so any speedup relies entirely on dropping from 5 steps to 3. However, the IL metric differences are small and lack multi-seed variance—which the appendix itself notes is close to NAVSIM evaluation noise. Meanwhile, Table 1 evaluates the Flow-GRPO baseline at against Engram at , making it impossible to tell whether the reported RL gains come from the memory module or simply from changing the inference step count.
Questions for authors
-
Please clarify the exact training procedure in Algorithm 2 versus Eq. 27: how are time steps sampled and integrated during the -step training rollout, are gradients backpropagated through the rollout states, and why is a valid supervision target off the linear interpolation path? How does a standard interpolation-trained Flow Matching baseline perform at and ?
-
What are the Flow-GRPO results when the baseline and Engram are evaluated at the exact same step count ( vs. , and vs. ), and what is the multi-seed variance for the main IL and RL comparisons?
4 comments
Reply to Reviewer WUiX: Comment 2, controlled performance and efficiency (W2 and Q2)
Thank you for recognizing the reduced-step planning result and the associated FLOP and latency measurements. Your concern is whether these comparisons isolate the memory contribution and establish a practical latency benefit. We distinguish matched-budget quality gains from the measured latency comparison below.
- Matched-budget RL (Q2). Both models were trained with K=5 rollouts. Evaluating the trained checkpoints without retraining gives the following navtest-v2 EPDMS scores. Engram-RL improves EPDMS by 3.94 points at K=3 and 4.04 at K=5. At K=3, NC improves from 94.08 to 98.37 and TTC from 92.99 to 97.79, despite lower EP (87.38 versus 93.04). Thus, the aggregate gain accompanies better safety-related metrics, not greater ego progress, and persists after controlling the step count. Both models are close to their five-step scores at K=3. The low-budget distinction is clearer at K=2: baseline-RL falls from 84.96 to 81.57, whereas Engram-RL changes from 88.90 to 88.78. This supports stronger quality retention at low step counts.
| Model | K | NC | DAC | DDC | TLC | EP | TTC | LK | HC | EC | EPDMS |
|---|---|---|---|---|---|---|---|---|---|---|---|
| baseline-RL | 1 | 94.72 | 92.48 | 98.01 | 100.00 | 90.07 | 95.64 | 96.32 | 97.08 | 40.13 | 77.17 |
| Engram-RL | 1 | 98.46 | 95.07 | 99.60 | 100.00 | 84.91 | 97.65 | 96.59 | 98.08 | 74.73 | 85.64 |
| baseline-RL | 2 | 91.68 | 95.05 | 98.64 | 100.00 | 94.54 | 90.60 | 94.81 | 97.95 | 80.16 | 81.57 |
| Engram-RL | 2 | 98.32 | 96.70 | 99.53 | 100.00 | 87.64 | 97.73 | 96.96 | 98.31 | 84.79 | 88.78 |
| baseline-RL | 3 | 94.08 | 96.06 | 98.87 | 100.00 | 93.04 | 92.99 | 95.20 | 98.30 | 84.73 | 84.96 |
| Engram-RL | 3 | 98.37 | 96.90 | 99.50 | 100.00 | 87.38 | 97.79 | 97.07 | 98.31 | 84.33 | 88.90 |
| baseline-RL | 5 | 94.42 | 96.03 | 98.86 | 100.00 | 92.80 | 93.27 | 94.98 | 98.33 | 84.37 | 85.02 |
| Engram-RL | 5 | 98.49 | 97.01 | 99.49 | 100.00 | 86.96 | 97.98 | 97.30 | 98.31 | 84.21 | 89.06 |
-
Planner quality and measured efficiency (W2). Structured action memory supplies reusable motion and refinement information, and the matched-budget RL results above show its planning-quality contribution. Separately, the latency measurements of IL provide a practical latency result: three-step Engram achieves 86.64 PDMS versus 86.19 for the five-step baseline, with 14.3% lower planner latency (72.3 versus 84.4 ms), including memory overhead. At K=3, the baseline scores 86.01 PDMS and is faster at 54.3 ms, so this saving applies to the reported configuration pair, not acceleration at equal K. The end-to-end reduction is more modest, approximately 4.8%, because the VLM dominates runtime. We will make the planner-level scope and comparison settings clearer, distinguishing the matched-budget quality gains from the measured latency saving.
-
Memory overhead (W2). The absolute planner timings above were measured on an idle GPU with CUDA synchronization and 50 measurements. Separate profiling at K=3 under concurrent GPU load estimates in table. CPU–GPU synchronization and small-kernel launches are the main identified bottlenecks. These percentages describe the separate profiling run, not a split of the idle-GPU timings. We will document both setups in App. 2.5.
| Component | Share of memory overhead |
|---|---|
| Hashing and lookup | 46% |
| Gated injection | 54% |
- Statistical reliability (W2/Q2). The reported scores and 95% scenario-bootstrap CIs are shown in table. The IL point estimates differ by 0.45 PDMS and have overlapping intervals; we describe this result as comparable observed quality, rather than a statistically established improvement. For RL, the 3.88-point gap between three-step Engram and the five-step baseline is accompanied by clearly separated intervals, supporting the reported difference with respect to scenario-sampling uncertainty on 12,146 scenarios. The matched-budget evaluations above address the step-count confound separately: gains remain at 3.94 and 4.04 points at K=3 and K=5. Thus, the RL quality benefit persists after controlling K, while the CIs here describe the original cross-budget comparison.
| Setting | Model | K | Score | 95% CI |
|---|---|---|---|---|
| IL, navtest-v1 (PDMS) | Engram | 3 | 86.64 | [86.23, 87.05] |
| IL, navtest-v1 (PDMS) | Baseline | 5 | 86.19 | [85.77, 86.63] |
| RL, navtest-v2 (EPDMS) | Engram-RL | 3 | 88.90 | [88.51, 89.27] |
| RL, navtest-v2 (EPDMS) | baseline-RL | 5 | 85.02 | [84.51, 85.56] |
We hope the matched-budget results, overhead analysis, and clarified statistical scope help assess the structured-memory contribution and would appreciate reconsideration of the assessment.
Official Comment by Reviewer
Thank you for your response. My concerns also remain unresolved.
Thanks for your response
Thank you very much for taking the time to review our responses and additional experiments. We understand that your concerns remain. We would appreciate any further clarification on which specific points our response has not sufficiently addressed, as this would help us provide a more focused follow-up.
Thanks again!
Reply to Reviewer WUiX: Comment 1, training procedure (W1 and Q1)
We thank the reviewer for recognizing the concrete memory architecture and the systematic IL ablations of capacity, branch selection, conditioning, and injection depth. Your question concerns whether the constant velocity target remains valid on the training states. The ambiguity comes from the term “rollout” in Sec. 3.7 and Alg. 2: during IL, it denotes sequential memory queries, not numerical integration of predicted velocities.
For each shared ($ ε, x^* $) pair, we draw K=3 independent time points from Beta(1.5, 1.0), map them through the noise schedule, and sort them in ascending order. At each ts, the query state is xts=(1−ts)ε+ts⋅x∗. All IL states therefore lie on the linear interpolation path; none is a self-integrated, off-path state. Continuous time sampling specifies these query times, while sorting establishes the noisy-to-clean order. The uniform integration step Δt=1/K is used only in inference and RL.
At every query, v∗=x∗−ε is the valid conditional FM target for this path. We compute the velocity MSE at each state and average the K losses with equal weight (sum/K). This is an average of conditional FM regression terms sharing a noise–target pair, with the history-dependent memory context supplied by CrossStep.
The states do not depend on predicted velocities. Predictions are detached before being written to the CrossStep buffer, so gradients do not propagate through earlier predictions or a state-integration chain. Each current prediction and the retrieved table entries receive gradients from the corresponding planning loss; the discrete hash indices do not. Thus, the procedure does not require full or truncated BPTT across query steps.
Sequential queries supply CrossStep with model-predicted velocity prefixes, the input type it uses at inference. This avoids constructing training keys from ground-truth velocities while using predicted velocities at test time. Training still queries interpolation states, whereas inference queries self-integrated states; this distinction does not invalidate the on-path IL supervision. We will make the state construction, time sampling, and detach operations explicit in Alg. 2 and App. 1.2.
To distinguish the memory contribution from this protocol, the memory-free baseline, Equal-FFN control, and branch variants all use the same training procedure. A standard random-time interpolation-trained baseline without sequential queries achieves 85.78 PDMS at K=3 and 86.01 at K=5 on navtest-v1, compared with 86.01 and 86.19 for the sequential-query baseline. The protocol comparison shows close single-run scores; the shared-protocol controls provide the direct comparison of memory architectures.
The addressing operations are fixed at initialization, while the embedding contents are learned through the planning loss. Engram therefore combines fixed symbolic addressing with learned contents. We hope these details resolve the mathematical and implementation ambiguity.