{"id":"28c24e6d-662f-4c4e-9459-9c83040beeb9","arxiv_id":"2505.21734","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A modular stack of a CNN localizer and a Stanley controller is trained end-to-end, and the upstream localizer learns to reduce downstream cross-track error.","lead":"MIND-Stack combines a lidar-based localization neural network with a classic Stanley steering controller into one differentiable pipeline, and shows the localization network can be trained to reduce the vehicle's path-following error. The result is a concrete recipe for getting end-to-end gradient training while keeping the autonomy stack modular and human-interpretable, which matters because safety-critical systems rarely trust full black-box networks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim lacks a matched training-duration control: reported before/after gains could be from extra gradient steps, not from end-to-end control loss.","rationale":"The reader's weakest_assumption focused on the unaltered localization module's bias in the training loss (Section III-C). That is a real concern, but it is not the most load-bearing for the central claim: even if the unaltered pose were perfect, the before/after design still fails to attribute the improvement to the downstream control loss rather than to additional training. The reader's rationale does mention 'the absence of a matched training-duration baseline,' so there is partial agreement, but it was not listed as the weakest assumption. I sharpened it into the primary stress-test because it directly targets the paper's strongest claim: that upstream localization reduces downstream control error through end-to-end differentiability. The open-source release and real-world deployment are genuine supporting evidence, but they do not remove the confound. The proposed concrete test is cheap and decisive: train a same-duration localization-only control and compare validation CTE. If the control matches the reported reductions, the headline result is an artifact of training budget; if not, the claim survives. Since the reader already issued CONDITIONAL and this concern can be resolved with an additional experiment, the verdict should remain CONDITIONAL, i.e., UNCHANGED from the reader's recommendation.","tokens_in":12193,"tokens_out":6504,"duration_ms":70151,"concrete_test":"Re-run the Scenario 1 and Scenario 5 protocols with a matched control: after the same 50-epoch Lloc pretraining, train the localization network for the same 30 laps, same optimizer and per-scenario learning rate, using only Lloc (Eq. 1), with no propagation through the Stanley controller or kinematic bicycle model. Then evaluate over 30 laps with the same validation metric. If this control reduces mean absolute cross-track error by the same margin as the reported 9.99% (Scenario 1) or 77.43% (Scenario 5), the control-loss gradient is not the cause; if the control shows no significant reduction while the end-to-end run does, the central claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline evidence (Table II: 9.99% to 77.43% validation-loss reductions) compares a localization network trained for 50 epochs on Lloc only ('Before Optimization') with the same network after an additional 30 laps of training on the control loss ('After Optimization'). The 'Localization Only' baseline in Table V is explicitly the before-optimization state, not a network given the same 30 additional laps of gradient updates on Lloc. Any additional training, even on the original localization objective, could improve pose estimates and therefore reduce path-tracking error; the unique contribution of the downstream control loss is not isolated. The same unmatched-budget problem affects the Section V-D comparisons: 'Stanley Controller Only' and 'Combined Training' are not shown to have identical training durations or update counts. Without a matched-duration Lloc-trained control, the central claim that end-to-end differentiability specifically enables the localization module to reduce downstream control error is not yet established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces MIND-Stack, a modular and fully differentiable autonomous-navigation stack that couples a CNN-based LiDAR localization module with a hand-crafted Stanley lateral controller. The localization network is first trained supervised on a localization loss, then fine-tuned on a combined control-oriented loss that includes cross-track error, an orientation-jerk penalty, and (in one scenario) the localization loss. Experiments in six simulation scenarios report 9.99%–77.43% reductions in mean absolute cross-track error after optimization, a comparison against several state-of-the-art controllers in one scenario, transfer of a simulation-trained model to a real F1Tenth vehicle with reduced errors, and a combined training setting in which both the localization network and the Stanley parameters are optimized jointly. The authors argue that this demonstrates a novel capability: an upstream localization module can reduce downstream control error through end-to-end differentiability while retaining modularity and interpretability.","tokens_in":12375,"tokens_out":6731,"duration_ms":69142,"significance":"If the central claim is sustained, MIND-Stack is a useful and timely contribution to hybrid modular autonomous driving: it spans from raw LiDAR input to steering output, is fully differentiable, keeps human-interpretable intermediate states, and is validated on an embedded real vehicle. The open-source code, the consistent per-timestep evaluation metrics, and the real-world deployment are explicit strengths that support reproducibility and practical relevance. However, the key causal claim—that end-to-end differentiability specifically enables the localization module to reduce downstream control error—is currently supported mainly by a before/after comparison that does not control for the amount of additional training the localization network receives. The absence of a matched training-duration baseline, together with the unquantified bias of the reference localization module used in the control-loss computation, leaves the central mechanism less certain than the abstract suggests. These are fixable with additional experiments, so the result is promising but not yet conclusive.","major_comments":[{"comment":"The main before/after comparison lacks a matched training-budget control. 'Before Optimization' is a localization network trained for 50 epochs on Lloc, while 'After Optimization' is the same network after an additional 30 laps of training on the control loss. There is no baseline that trains the localization network for the same 30 additional laps on Lloc alone. Any additional gradient updates, even if supervised by the original localization loss, could improve pose estimates and hence reduce the ground-truth cross-track error reported as validation loss. The same issue appears in the 'Localization Only' row of Table V and in the real-world transfer comparison of Table IV, where the after-optimization model has simply received more training than the before-optimization model. Please add a control that trains the localization network for the same number of laps/updates on Lloc only and report its validation loss, so the specific contribution of the downstream control loss is isolated.","section":"§V-A, Table II; §V-D, Table V"},{"comment":"The cross-track error term in Eq. (3) is computed from a pose propagated from an 'unaltered localization module' running in parallel with the trained network. The paper states that this module provides 'an accurate pose estimate,' but it never quantifies the module's own error or systematic bias relative to ground truth in the simulation or real-world scenarios. If the reference pose has a systematic offset, the control loss is optimized against a biased error signal, and the observed reductions in validation loss could be a side effect of the network compensating for that bias rather than evidence that it improves true path-following. Please report the unaltered module's localization error/bias in each scenario and, ideally, ablate the cross-track error computation by using the ground-truth pose as the reference.","section":"§III-C"},{"comment":"In most scenarios γ is absent (Table I), so the end-to-end loss does not constrain the localization output to the ground-truth pose. The paper claims interpretability through intermediate pose representations, but it never measures how far the optimized localization output deviates from the unaltered module or from ground truth. If the network learns a pose that is far from the true pose while producing low control error, the 'human interpretable state representation' is no longer an accurate vehicle state, weakening a stated contribution. Please report the localization error of the optimized network and explicitly discuss the trade-off between control loss and pose interpretability.","section":"§V-D, Table I"}],"minor_comments":[{"comment":"The orientation error term in Eq. (1) uses the raw difference (θ̂ − θ) without wrapping to the interval [−π, π]; this can produce discontinuous gradients near ±π and may distort the localization loss in scenarios where the heading crosses that boundary.","section":"Eq. (1)"},{"comment":"The '–' entries for γ in Table I are not explained in the text; please state explicitly that those scenarios set γ = 0 (or otherwise omit the localization term from Ltotal).","section":"Table I"},{"comment":"The real-world experiments operate at a 40 Hz control rate while the simulation uses 100 Hz, but the paper does not state whether the simulation-trained model and the kinematic bicycle model in the loss were trained at 100 Hz and how the transfer to 40 Hz affects the two-step propagation time step; please clarify.","section":"§V-C"},{"comment":"The state-of-the-art comparison is reported for a single scenario without a statistical significance test; adding a paired test across the 30 evaluation laps would strengthen the claim that MIND-Stack outperforms the listed controllers.","section":"Table III"},{"comment":"There are minor language issues, including 'in-, output' in the Related Work section and 'at competitive speeds and lap times' in Section V-B; a careful proofread would improve clarity.","section":"§I, Related Work"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid system demonstration with open-source code and real-world validation, but the central scientific claim about end-to-end differentiability is currently undercut by the missing matched-training-duration control. I would be willing to accept a revision that adds (1) a localization-only training control with the same number of additional laps, (2) an error/bias characterization of the unaltered reference module, and (3) an analysis of the optimized network's pose deviation to support the interpretability claim. These are empirical additions within the existing framework rather than fundamental changes."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read MIND-Stack. The core idea is sensible: string a CNN localizer, a differentiable Stanley controller, and a kinematic bicycle model into one graph, then train the localizer on a downstream control loss. The individual pieces are known, but the specific combination—sensor-to-actuator differentiability with a hand-crafted control law and real F1Tenth deployment—is not something I've seen in the prior work they cite. The paper deserves credit for shipping open-source code and for showing consistent validation-loss improvements across six simulation scenarios plus a real-world sim-to-real transfer. The simultaneous training of the localizer and Stanley gains (Table V) is a nice demonstration of the architecture's flexibility.\n\nThe main weakness is the missing training-duration control. The \"before\" state is a network trained for 50 epochs on Lloc only; the \"after\" state is that same network plus 30 laps of additional gradient updates on the control loss. There is no baseline that gets the same 30 additional laps of updates on Lloc alone. Any extra training, even on the original objective, could tighten the pose estimates and lower the path-tracking error. So the headline claim—that end-to-end differentiability specifically trains the localizer to reduce downstream control error—is not yet isolated. The same issue affects Table V: the four setups are not shown to have matched update counts. Adding that control is straightforward and would make the central result much cleaner.\n\nA second, smaller soft spot: the training loss's cross-track error is computed using a parallel \"unaltered\" localization module, and the paper never quantifies that module's own error or bias. If that module has a systematic offset, the loss could be training against the wrong error signal. This is minor, but worth addressing.\n\nAlso, the \"first fully differentiable stack\" phrasing is a bit strong; the paper does enough to distinguish itself from DiffStack, DAN, and PyPose, but that claim should be softened to \"to our knowledge\" and perhaps with more precise bounds on prior art. The SOTA comparison runs on only one scenario, so the \"outperforms SOTA\" claim rests on a thin base, though the baseline already being strong is a point in their favor.\n\nOverall, this is a solid engineering paper with an addressable flaw. It would benefit from a matched-budget ablation and a quantification of the parallel localizer's accuracy. I'd send it to referees; it's the kind of work where a careful revision could make the central claim stick.","headline":"Solid engineering paper with an addressable flaw: the headline end-to-end training benefit is not isolated from extra training time, but the architecture and real-world demo are worth refereeing.","tokens_in":12885,"tokens_out":4368,"would_cite":false,"duration_ms":39144,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that training a localization network on the downstream control loss of a differentiable Stanley controller reduces path-following error by up to 77.43% in simulation and 50.77% on a real vehicle.","keywords":["autonomous navigation","end-to-end differentiability","modular software stack","localization","Stanley controller","cross-track error","sim-to-real transfer","embedded systems"],"falsifier":"Measure the unaltered parallel localization module's own mean absolute cross-track error in each scenario; if it is comparable to or larger than the reported 9.99-77.43% reductions, the control loss may be fitting to that module's bias rather than to true path-tracking. A direct way to test this is to retrain in simulation using ground-truth pose to compute the cross-track error and see whether the gains persist.","tokens_in":11993,"feed_emoji":"🚗","tokens_out":13283,"duration_ms":118793,"temperature":0.7,"pith_summary":"MIND-Stack is a modular autonomous-navigation stack that keeps a conventional rule-based controller—a Stanley steering law—but makes the whole pipeline from LiDAR to steering angle differentiable. The paper tries to establish that the upstream localization network can be trained on the downstream control loss (a squared cross-track error), so that the network's job is not merely to estimate the vehicle's pose but to produce pose estimates that reduce path-following error. In simulation across six track segments, this reduces the mean absolute cross-track error by 9.99% to 77.43%, with an average reduction of 45.61%, and the improvement transfers to a real vehicle with up to a 50.77% reduction. The authors also show that training the localization network and the controller's hand-tuned gains simultaneously outperforms training either module alone. If the result holds, it would mean that modular, interpretable autonomy stacks do not have to give up end-to-end learning.","feed_headline":"Localization trained on path error cuts tracking error up to 77%","feed_subtitle":"A differentiable sensor-to-steering stack lets the localization module learn from path error, not just pose error.","key_machinery":"The central mechanism is the differentiable Stanley controller, a steering law $\\delta = k_h e_{\\mathrm{head}} + \\arctan(k_e e_{\\mathrm{cross}}/v)$ that maps a pose estimate to a steering angle and is differentiable with respect to that pose. To make the control loss flow into the network, the stack propagates the network's pose estimate two timesteps (0.01 s each) through a kinematic bicycle model, so the cross-track error $e_{\\mathrm{cross}}$ depends on the steering command and therefore on the pose. A parallel unaltered localization module, not the trained network, provides the pose used to compute that cross-track error, which preserves interpretability of the network's output. The training objective is $L_{\\mathrm{total}} = \\alpha e_{\\mathrm{cross}}^2 + \\beta|\\theta_t - 2\\theta_{t-1} + \\theta_{t-2}| + \\gamma L_{\\mathrm{loc}}$, mixing the downstream error, a heading-smoothness term, and the original localization loss.","core_discovery":"The paper's central claim is that end-to-end differentiability through a modular stack lets an upstream localization module be optimized for the downstream control objective, not just for its own supervised loss. Concretely, the localization CNN is first trained on a pose-reconstruction loss, then fine-tuned on a total loss that combines a squared cross-track error (computed by propagating the trained network's pose two timesteps through a kinematic bicycle model), a smoothness penalty on heading, and the original localization loss. Gradient flow is made possible because the Stanley controller's steering formula is differentiable with respect to its pose input. In six simulated scenarios the validation loss, mean absolute cross-track error per timestep relative to ground truth, falls by between 9.99% and 77.43%; in two real-world scenarios it falls by up to 50.77%. The paper interprets this as evidence that a localization module can be trained to serve the controller, and that combined training of the controller's gains and the network yields a further 12% validation-loss reduction over the second-best setup.","pith_inferences":["A likely mechanism behind the reported reductions is that control-loss training adjusts the pose estimate to compensate for the fixed controller's steady-state errors, so the network's output becomes less a pure sensor estimate and more a controller-aware correction; one could test this by comparing the trained network's pose output with the untrained module's output and with ground truth.","The general recipe suggested here is that any differentiable downstream cost can fine-tune upstream modules, so substituting a planning or MPC cost for the cross-track error could shape different driving behaviors, such as aggressiveness or comfort.","Because the cross-track error in training is computed relative to the parallel unaltered localization module, the reported improvements could partly be the network learning to reproduce that module's bias rather than learning better geometry; a direct measurement of that module's own error would bound the effect.","The scenario spread (9.99% minimum, 77.43% maximum) suggests the benefit is environment-dependent, so future work could characterize when control-loss training helps most—for example, on tracks where the baseline localization is weak relative to the controller."],"forward_implications":["A localization network trained on the control loss can cut mean absolute cross-track error by up to 77.43% in simulation and 50.77% on a real vehicle, without changing the controller.","The same differentiable-stack design can be extended to perception, prediction, and planning modules so each is optimized for the final driving objective rather than only its local loss.","Simultaneous training of a neural module and a hand-crafted controller's parameters is feasible and yields lower path error than training either module alone.","The pipeline runs in about 0.019 s per inference on an embedded GPU board, faster than the 40 Hz LiDAR input, so it is deployable on low-power hardware."],"supporting_citations":[{"why":"Prior differentiable control stack that motivates the end-to-end differentiability idea but lacks sensor-to-actuator coverage and real-world validation.","marker":"[6]"},{"why":"Defines the Stanley steering law used as the differentiable control module and as the basis for the cross-track error loss.","marker":"[34]"},{"why":"Supplies the kinematic bicycle model that propagates the pose estimate two timesteps so the control error depends on the steering output.","marker":"[39]"},{"why":"Justifies generating the previous pose during training by adding noise around ground truth, which is how the localization training data are produced.","marker":"[38]"},{"why":"Follow-the-Gap baseline algorithm against which MIND-Stack's path-following error is compared.","marker":"[42]"},{"why":"Pure Pursuit baseline controller used in the state-of-the-art comparison.","marker":"[43]"},{"why":"Provides the particle-filter localization used with the Pure Pursuit, Stanley, and MPC baselines in the comparison.","marker":"[44]"},{"why":"Supplies the small-scale vehicle platform and simulator used for the real-world and simulated experiments.","marker":"[40]"}],"fun_headline_variants":["Localization tuned on control error cuts cross-track error up to 77%","MIND-Stack: Differentiable modular stack reduces tracking error by 77%","End-to-end differentiable autonomy: localization learns from steering","Modular, interpretable stack with end-to-end training cuts path error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The unaltered parallel localization module's pose is assumed accurate enough to serve as the reference for the cross-track error in the training loss, but the paper never quantifies that module's own error or systematic bias.","fun_headline_variants_meta":{"raw":{"variants":["Localization tuned on control error cuts cross-track error up to 77%","MIND-Stack: Differentiable modular stack reduces tracking error by 77%","End-to-end differentiable autonomy: localization learns from steering","Modular, interpretable stack with end-to-end training cuts path error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000819,"raw_usage":{"total_tokens":3604,"prompt_tokens":980,"completion_tokens":2624,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":2546}},"tokens_in":596,"tokens_out":2624,"duration_ms":21393,"temperature":1.0,"reasoning_tokens":2546,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:23:57.514540+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the unaltered parallel localization module's own mean absolute cross-track error in each scenario; if it is comparable to or larger than the reported 9.99-77.43% reductions, the control loss may be fitting to that module's bias rather than to true path-tracking. A direct way to test this is to retrain in simulation using ground-truth pose to compute the cross-track error and see whether the gains persist.","supporting_citations":[{"cited_title":"Diffstack: A differentiable and modular control stack for autonomous vehicles,","cited_arxiv_id":null,"evidence_quote":"Prior differentiable control stack that motivates the end-to-end differentiability idea but lacks sensor-to-actuator coverage and real-world validation."},{"cited_title":"Autonomous Automobile Trajectory Tracking for Off-Road Driving: Controller Design, Experimental Validation and Racing,","cited_arxiv_id":null,"evidence_quote":"Defines the Stanley steering law used as the differentiable control module and as the basis for the cross-track error loss."},{"cited_title":"CommonRoad: Com- posable benchmarks for motion planning on roads,","cited_arxiv_id":null,"evidence_quote":"Supplies the kinematic bicycle model that propagates the pose estimate two timesteps so the control error depends on the steering output."},{"cited_title":"Local INN: Implicit map representation and localization with invertible neural networks,","cited_arxiv_id":null,"evidence_quote":"Justifies generating the previous pose during training by adding noise around ground truth, which is how the localization training data are produced."},{"cited_title":"A novel obstacle avoidance algorithm: “Follow the Gap Method","cited_arxiv_id":null,"evidence_quote":"Follow-the-Gap baseline algorithm against which MIND-Stack's path-following error is compared."},{"cited_title":"Implementation of the pure pursuit path tracking algorithm,","cited_arxiv_id":null,"evidence_quote":"Pure Pursuit baseline controller used in the state-of-the-art comparison."},{"cited_title":"Cddt: Fast approximate 2d ray casting for accelerated localization,","cited_arxiv_id":null,"evidence_quote":"Provides the particle-filter localization used with the Pure Pursuit, Stanley, and MPC baselines in the comparison."},{"cited_title":"F1TENTH: An Open-source Evaluation Environment for Continuous Control and Reinforcement Learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the small-scale vehicle platform and simulator used for the real-world and simulated experiments."}],"review_version":1}