{"id":"54ece4c6-475c-4910-b969-cec64b7a94a8","arxiv_id":"2411.15139","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"DiffusionDrive shows that seeding a diffusion policy with K-Means anchor trajectories and truncating the diffusion schedule allows real-time (45 FPS) end-to-end driving planning with a 2-step denoising process and a 88.1 PDMS score on NAVSIM.","lead":"DiffusionDrive is a self-driving planning model that starts its diffusion process from a small set of typical driving trajectories instead of pure noise, so it needs only 2 refinement steps instead of 20. On the NAVSIM benchmark it reports 88.1 PDMS, a new state of the art, while running at 45 frames per second on an NVIDIA 4090.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 88.1 PDMS result depends on 20 training-set anchors covering all test-time driving modes; the paper offers no quantitative coverage analysis, only qualitative examples and a CARLA transfer test.","rationale":"The reader's weakest_assumption identifies anchor coverage as the key risk, and I agree. The paper's core novelty is the truncated diffusion policy starting from a small set of learned anchors. If those anchors do not span the action space encountered at test time, the method inherits exactly the out-of-vocabulary failure it criticizes in VADv2, and its headline PDMS on navtest may not reflect robustness in more varied or adversarial scenarios. The paper provides only qualitative illustrations and a CARLA transfer experiment, neither of which measures coverage on navtest. My proposed test is a direct quantitative check: compare PDMS for navtest ground truths that are far from the nearest anchor versus those that are near, and ablate the anchor count. This targets the assumption without requiring new data collection. The reader's CONDITIONAL verdict remains appropriate because the concern is addressable with additional analysis; it does not invalidate the existing results outright, but it does mean the diversity and open-world generalization claims are not yet fully supported.","tokens_in":16353,"tokens_out":9416,"duration_ms":88697,"concrete_test":"Compute the Euclidean distance (after the normalization used in K-Means) between each navtest ground-truth trajectory and its nearest of the 20 anchors. Split navtest into quintiles by this distance; if the farthest quintile's PDMS is more than 5 points below the nearest quintile's, the 20-anchor coverage is insufficient and the 88.1 average is masked by easy scenes. Additionally, rerun K-Means with 10, 50, and 100 anchors on the same training set and retrain/fine-tune DiffusionDrive; if PDMS at 20 anchors is far below the 100-anchor result, the anchor count is a critical bottleneck rather than a robust design choice.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a 2-step truncated diffusion planner achieves 88.1 PDMS on NAVSIM navtest and, more broadly, generates diverse multi-mode actions. Inference starts from an anchored Gaussian distribution built from only 20 K-Means clusters of the training set (Sec. 3.3, Eq. 4, T_trunc=50/1000). Since alpha_bar_50 is close to 1, the starting noise is small, so sampled trajectories are tightly concentrated around the 20 anchors; the 2 DDIM steps provide limited ability to move far from these anchors. The model can therefore only produce trajectories near the training-cluster modes unless the decoder extrapolates arbitrarily, but the training loss (Eq. 6) only supervises reconstruction of each ground-truth trajectory from its closest anchor, so the model is never trained to handle test-time scenarios whose required action is far from every anchor. The paper asserts that DiffusionDrive avoids the out-of-vocabulary failure of VADv2's 8192-anchor vocabulary, yet its only evidence for anchor coverage is qualitative (Fig. 2, supp. Figs. 5-7) and a CARLA transfer test (Table 9) that uses the same NAVSIM anchors on a different dataset, not a coverage analysis on navtest. If a navtest scene requires a mode in a gap between anchors, the top-1 trajectory could be arbitrarily poor, undermining the 88.1 PDMS as a generalizable result.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DiffusionDrive, an end-to-end driving planner that replaces the deterministic regression head of Transfuser with a truncated diffusion policy. The method uses K-Means anchors from the training set to define an anchored Gaussian distribution, truncates the diffusion schedule to T_trunc = 50/1000, and performs only 2 denoising steps at inference with a cascade transformer decoder. On the NAVSIM navtest split, DiffusionDrive achieves 88.1 PDMS with a ResNet-34 backbone at 45 FPS on an NVIDIA 4090, and it also reports improved L2 error and collision rate on nuScenes over SparseDrive. The paper includes a controlled roadmap from Transfuser to TransfuserDP to TransfuserTD to DiffusionDrive, ablations of decoder components, and a CARLA transfer experiment for anchor generalization.","tokens_in":16714,"tokens_out":5382,"duration_ms":50266,"significance":"If the central claim holds, the paper makes a meaningful contribution: it shows that a diffusion-based planner can be made real-time (2 steps) while achieving state-of-the-art closed-loop planning performance on NAVSIM and providing diverse multi-mode trajectories. The controlled roadmap and component ablations are strengths, and the CARLA anchor-transfer experiment is a useful step toward validating the generality of the anchor prior. The performance on an externally defined benchmark (NAVSIM navtest) is a significant asset. However, the core claims about avoiding out-of-vocabulary modes and about diversity rest partly on qualitative evidence and on a self-referential diversity score, which need to be substantiated with quantitative analysis.","major_comments":[{"comment":"The claim that the anchored Gaussian distribution with only 20 K-Means anchors avoids the out-of-vocabulary failure of VADv2's 8192-anchor vocabulary is not quantitatively supported. Since Eq. (4) with T_trunc=50/1000 yields a starting noise concentrated near the anchors, and the training loss in Eq. (6) reconstructs the ground truth only from its closest anchor, a navtest scenario requiring a mode far from every anchor could produce arbitrarily poor top-1 predictions. The CARLA transfer test (Table 9) uses the same NAVSIM anchors on a different dataset and does not measure coverage of navtest modes. Please add a coverage analysis on navtest, such as the distribution of distances from ground-truth trajectories to the nearest anchor, PDMS as a function of anchor count, or a failure-case study for scenarios where the top-1 trajectory deviates from the ground truth.","section":"Section 3.3 / Table 9"},{"comment":"The ablation in Table 3 does not include a non-diffusion baseline that uses the same cascade decoder and anchor classification but performs a single forward pass without iterative DDIM denoising. As a result, the improvement attributed to the truncated diffusion process could be due to the anchor prior, the classification head, or the transformer decoder itself. Please add such a baseline (e.g., training the decoder to directly regress offsets from anchors with no iterative refinement) to isolate the contribution of the diffusion/denoising mechanism.","section":"Section 3.4 / Table 3"},{"comment":"The mode diversity score D (Eq. 3) is a spatial-spread measure (1 minus the mean intersection-over-union of sampled trajectories). It is computed on the model's own outputs and is not validated against any notion of plausibility or multi-modality quality; a model that outputs arbitrary scattered trajectories could achieve a high D. Since the paper's central claim includes 'superior diversity and quality,' please validate D against an external criterion (e.g., human ratings, downstream closed-loop performance of the diverse modes, or comparison with trajectory prediction benchmarks) and report D for the non-diffusion baseline as well.","section":"Section 3.2 / Eq. (3)"}],"minor_comments":[{"comment":"The claim that this work 'firstly introduce[s] the diffusion model to the field of end-to-end autonomous driving' should be reconciled with the cited GenAD [58] and other generative planning works; if GenAD uses a diffusion-style generator, the novelty claim needs to be qualified.","section":"Section 1"},{"comment":"The column header 'D' is not explained in the caption; define the mode diversity score in the caption or refer explicitly to Eq. (3).","section":"Table 2"},{"comment":"The sentence 'resulting in a total 650x increase in runtime overhead' is correct only when comparing the 20-step UNet to the 1-step MLP; please state the comparison base explicitly.","section":"Section 4.4"},{"comment":"The caption contains a typo: 'DiffusioDrive' should be 'DiffusionDrive'.","section":"Supplementary Table 8"}],"recommendation":"major_revision","confidential_remarks":"The empirical results are strong and the roadmap/ablations are informative. The main risk is the anchor-coverage assumption, which the authors should be asked to quantify. Please also verify the originality claim against GenAD and other diffusion-based planners before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a serious, well-executed paper. The main claim—a truncated diffusion policy that starts from an anchored Gaussian prior can plan in two denoising steps and beat the previous NAVSIM SOTA by 1.6 PDMS without post-processing—is backed by a controlled roadmap (Transfuser → TransfuserDP → TransfuserTD → DiffusionDrive) and sensible ablations. The efficiency gain is real: 45 FPS on a 4090, 10× fewer denoising steps than vanilla diffusion policy.\n\nWhat’s new: the combination of K-means anchors with a truncated diffusion schedule is a genuine twist on TDPM/DDIM applied to driving policy, and the cascade decoder with deformable cross-attention is a reasonable architectural contribution. The nuScenes results, while open-loop, are competitive.\n\nThe soft spots are real but not deal-breakers. The biggest is anchor coverage. With T_trunc=50/1000, ᾱ is close to 1, so inference starts almost exactly on the 20 training anchors; the model has little room to escape that vocabulary. The paper’s defense is qualitative plus a CARLA transfer test, but there is no quantitative analysis of how often navtest ground-truth trajectories fall near an anchor. That leaves the possibility of out-of-vocabulary failure in edge cases. The fix is easy: report the distribution of distances from test trajectories to the nearest anchor, and possibly add a small noise level during training to increase robustness.\n\nSecond, there is no baseline with only the 20 anchors scored by the same decoder without diffusion; that would isolate whether diffusion actually adds value over a fixed-anchor scorer. Given that the paper touts 400× anchor reduction, this is a fair ask.\n\nThird, inference sampling from the anchored Gaussian is underspecified: exactly how are Ninfer noises drawn? One per anchor? A mixture with noise level from T_trunc? The paper says 'sampled from anchored Gaussian distribution' but doesn't give the formula for inference. The mode diversity score D is also self-referential—it measures spread, not correctness of modes.\n\nFinally, code and checkpoints are not yet public, which makes verification hard. The paper promises them, but at review time they are missing.\n\nOverall, this is a good paper for the end-to-end driving community. It demonstrates a practical way to make diffusion planning fast. The limitations are addressable and typical for the field. I’d send this to peer review and ask for the anchor coverage analysis, the non-diffusion baseline, and code release.","headline":"Real 2-step diffusion planner with anchored Gaussian prior; strong engineering, but anchor coverage and inference sampling need tightening.","tokens_in":17230,"tokens_out":2899,"would_cite":true,"duration_ms":27907,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A diffusion planner that denoises from 20 clustered anchors instead of pure noise reaches 88.1 PDMS on NAVSIM in 2 steps, at 45 FPS.","keywords":["truncated diffusion policy","anchored Gaussian distribution","end-to-end autonomous driving","multi-mode planning","cascade diffusion decoder","NAVSIM","denoising steps","real-time planning"],"falsifier":"Run DiffusionDrive on a set of NAVSIM navtest scenarios deliberately chosen to contain maneuvers rare or absent in the training clusters — sharp u-turns, emergency swerves, unusual intersection geometry — and compare PDMS with the reported 88.1; a large drop on those scenes would show the anchored-coverage assumption fails. A cheaper check is to measure, for each denoised trajectory, the distance to the nearest training anchor and correlate high-distance samples with collisions or drivable-area violations.","tokens_in":16172,"feed_emoji":"🚗","tokens_out":10170,"duration_ms":78699,"temperature":0.7,"pith_summary":"The paper tries to establish that a probabilistic diffusion planner can meet the real-time and multi-mode demands of end-to-end driving, provided denoising starts not from pure Gaussian noise but from an 'anchored Gaussian distribution' built around a small set of clustered trajectory priors. This truncated diffusion policy cuts the required denoising steps from twenty to two, a tenfold reduction, while producing trajectories that are more diverse and more accurate than those of a vanilla diffusion head. On the NAVSIM planning benchmark, with the same ResNet-34 backbone as the Transfuser baseline, DiffusionDrive reaches 88.1 PDMS at 45 FPS on an NVIDIA 4090, beating the prior best 86.5 PDMS scored by a system that uses 8,192 anchors, rule-based supervision, and post-processing. A sympathetic reader would care because the result suggests generative multi-mode planning can run in real time and outperform carefully built sampling-from-vocabulary pipelines without post-processing.","feed_headline":"Diffusion planner hits 88.1 PDMS in just 2 denoising steps","feed_subtitle":"A 20-anchor truncated diffusion policy outruns an 8,192-anchor vocabulary rival at 45 FPS on NAVSIM.","key_machinery":"The load-bearing object is the truncated diffusion policy: the forward process in Eq. (4), $\\tau^i_k = \\sqrt{\\bar{\\alpha}_i}\\, a_k + \\sqrt{1-\\bar{\\alpha}_i}\\,\\epsilon$, diffuses each K-Means anchor trajectory $a_k$ only up to timestep $T_{\\text{trunc}}=50$ of the 1000-step schedule, so training and inference work on a mixture of sub-Gaussians around the anchors rather than on pure noise. The second component is the cascade diffusion decoder, a transformer layer that applies deformable spatial cross-attention to BEV or PV features, cross-attends to agent and map queries from the perception module, modulates on the diffusion timestep, and predicts both a confidence score and a trajectory offset per sample; two such layers are stacked with parameters shared across denoising steps. Together they let a variable number of samples $N_{\\text{infer}}$ be drawn near the anchors, denoised in two steps with the DDIM update rule, and ranked by the learned confidence score so that the top-1 trajectory carries the final plan.","core_discovery":"The central claim is that the obstacle to diffusion-based driving is not the generative model itself but the starting distribution: vanilla policies that denoise from random Gaussian noise suffer mode collapse in traffic scenes and need about twenty DDIM steps, dropping frame rate from 60 to 7 FPS. DiffusionDrive instead clusters the training trajectories into 20 anchors via K-Means, truncates the diffusion schedule to the first 50 of 1000 steps so each anchor is perturbed only into a small surrounding sub-Gaussian, and learns to denoise from that anchored distribution in 2 steps. The paper argues that this combines the coverage of a generative model with the structure of human driving patterns, so 20 anchors replace the fixed vocabulary of 8,192 anchors used by VADv2, and the top-1 scored denoised trajectory achieves 88.1 PDMS on the NAVSIM navtest split, 4.1 points above the Transfuser regression baseline and 1.6 points above Hydra-MDP-V8192-W-EP, with the full sampled set showing diverse maneuvers such as lane changes and traffic-light-aware stopping.","pith_inferences":["The twenty-cluster premise invites a stress test the paper does not run: a curated set of test scenes whose ground-truth maneuvers lie far from every training anchor should expose the coverage boundary, and if PDMS degrades sharply there, the method inherits the same out-of-vocabulary failure it attributes to VADv2.","The truncated-schedule recipe is generic: any policy task where a modest set of prior modes can be derived cheaply from data could start denoising from an anchored distribution, though the benefit should shrink as mode coverage becomes harder to certify with a small cluster count.","The diversity score $D$ of Eq. (3) measures trajectory spread, not usefulness; a natural extension is to test whether the diverse top-10 trajectories improve closed-loop robustness under reactive agents, which the top-1-based PDMS metric does not capture.","Because one-step inference already scores 87.9 PDMS, a dynamic step-count scheduler conditioned on scene difficulty is a plausible extension: easy scenes run one denoising step and hard scenes two or three."],"forward_implications":["A two-step denoising loop at 45 FPS makes generative multi-mode planning viable in real-time on-vehicle systems, where vanilla diffusion's twenty steps (7 FPS) were impractical.","Replacing a fixed 8,192-anchor vocabulary with 20 learned anchors plus continuous denoising removes the out-of-vocabulary failure the paper attributes to sampling-from-vocabulary planners, while cutting the anchor set 400-fold.","The cascade diffusion decoder improves PDMS by 2.4 over the UNet-based truncated policy while reducing parameters from 102M to 60M, pointing to the interaction design as the source of planning quality.","On nuScenes open-loop evaluation the same mechanism lowers average L2 error by 20.8% and collision rate by 63.6% relative to VAD with a ResNet-50 backbone, showing the result is not benchmark-specific.","Because sample count $N_{\\text{infer}}$ and the number of denoising steps are adjustable at inference, the system exposes a direct quality-versus-latency trade-off for deployment."],"supporting_citations":[{"why":"Baseline policy whose 20-step denoising DiffusionDrive replaces; supplies the vanilla diffusion head used for TransfuserDP.","marker":"[6]"},{"why":"The deterministic single-mode planner DiffusionDrive builds on; same perception modules and ResNet-34 backbone make the 4.1 PDMS gain attributable to the planning module.","marker":"[7]"},{"why":"The fixed-vocabulary paradigm (8,192 anchors) that the paper argues fails out-of-vocabulary; the anchor-number comparison baseline.","marker":"[3]"},{"why":"The previous state-of-the-art multi-mode planner whose W-EP variant (86.5 PDMS) is the main number DiffusionDrive beats by 1.6 PDMS.","marker":"[25]"},{"why":"Provides the NAVSIM dataset, the PDMS metric, and the navtest split used for the headline closed-loop benchmark.","marker":"[10]"},{"why":"The DDIM update rule used for the truncated denoising process during inference.","marker":"[35]"},{"why":"The ResNet-34 backbone shared across NAVSIM comparisons for an aligned benchmark.","marker":"[13]"},{"why":"Base implementation on nuScenes whose perception queries and PV features feed the diffusion decoder in the second benchmark.","marker":"[39]"}],"fun_headline_variants":["Truncated diffusion drives 2-step planning to 88.1 PDMS","20 anchors, 2 steps: DiffusionDrive hits 88.1 PDMS","DiffusionDrive slashes denoising steps 10x to hit 88.1","2-step truncated diffusion: self-driving record 88.1 PDMS","No random noise: anchored diffusion drives in 2 steps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that twenty trajectory clusters computed on the training set cover every driving mode the test scenarios will demand; a scene that needs a maneuver far from all anchors falls outside the model's generative range, reproducing the out-of-vocabulary failure the paper attributes to VADv2.","fun_headline_variants_meta":{"raw":{"variants":["Truncated diffusion drives 2-step planning to 88.1 PDMS","20 anchors, 2 steps: DiffusionDrive hits 88.1 PDMS","DiffusionDrive slashes denoising steps 10x to hit 88.1","2-step truncated diffusion: self-driving record 88.1 PDMS","No random noise: anchored diffusion drives in 2 steps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000702,"raw_usage":{"total_tokens":3209,"prompt_tokens":1027,"completion_tokens":2182,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":643,"completion_tokens_details":{"reasoning_tokens":2081}},"tokens_in":643,"tokens_out":2182,"duration_ms":15624,"temperature":1.0,"reasoning_tokens":2081,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:27:41.651139+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DiffusionDrive on a set of NAVSIM navtest scenarios deliberately chosen to contain maneuvers rare or absent in the training clusters — sharp u-turns, emergency swerves, unusual intersection geometry — and compare PDMS with the reported 88.1; a large drop on those scenes would show the anchored-coverage assumption fails. A cheaper check is to measure, for each denoised trajectory, the distance to the nearest training anchor and correlate high-distance samples with collisions or drivable-area violations.","supporting_citations":[{"cited_title":"Diffusion policy: Visuomotor policy learning via action dif- fusion","cited_arxiv_id":null,"evidence_quote":"Baseline policy whose 20-step denoising DiffusionDrive replaces; supplies the vanilla diffusion head used for TransfuserDP."},{"cited_title":"Transfuser: Imitation with transformer-based sensor fusion for autonomous driv- ing","cited_arxiv_id":null,"evidence_quote":"The deterministic single-mode planner DiffusionDrive builds on; same perception modules and ResNet-34 backbone make the 4.1 PDMS gain attributable to the planning module."},{"cited_title":"Navsim: Data-driven non- reactive autonomous vehicle simulation and benchmarking","cited_arxiv_id":null,"evidence_quote":"Provides the NAVSIM dataset, the PDMS metric, and the navtest split used for the headline closed-loop benchmark."},{"cited_title":"Denois- ing diffusion implicit models","cited_arxiv_id":null,"evidence_quote":"The DDIM update rule used for the truncated denoising process during inference."}],"review_version":1}