{"id":"bb9b7868-c503-47d6-9c39-7b711ec84750","arxiv_id":"2509.04213","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A transformer dynamics model plugged into an unscented Kalman filter gives zero-shot state estimates on unseen simulated container ships and outperforms classical baselines when velocity sensors are missing.","lead":"The paper combines a transformer-based model of ship dynamics with a Kalman filter to estimate the state of container ships it has never seen, without retraining for new sensor layouts. The result is an open-source benchmark and a method that beats classical approximations when velocities are not directly measured.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sigma-point context in FM-UKF is an unvalidated distribution shift; an oracle-context comparison is needed.","rationale":"The reader's weakest-assumption identification is spot-on. The paper's own wording in Sec. IV-A flags the assumption, but no experiment isolates it. The central claim relies on the transformer being an effective dynamics prior under sigma-point context; without validating that, the method is not yet trustworthy as a zero-shot tool. I find no more fundamental internal inconsistency: the benchmark is thoughtfully constructed, the comparisons are relevant, and the results do show the FM-UKF is competitive in the specified setting. Thus CONDITIONAL is the right verdict, and my stress-test does not move it.","tokens_in":10674,"tokens_out":4894,"duration_ms":45767,"concrete_test":"During FM-UKF evaluation on held-out ships, record the sigma-point context fed to the transformer at each step. In simulation, also feed the true state history (plus the same control inputs) through the same transformer to obtain an 'oracle-context' one-step prediction. Compare mean prediction error of the FM on sigma-point context vs true context for each feature. If the oracle-context error is substantially lower (e.g., >50% relative), the sigma-point assumption is violated and the method's success is contingent on an unvalidated distribution shift. Additionally, retrain the FM with training examples that mimic UKF posterior sigma points (e.g., add correlated noise), and check whether FM-UKF accuracy improves; if it does, the original training distribution was mismatched.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-A states the load-bearing assumption: 'the sequence of each n-th sigma point trajectory forms a coherent and sufficient context for the transformer.' The dynamics transformer is trained on true state histories (with small input noise), but during FM-UKF inference it receives histories of sigma points that have been updated by the UKF and therefore lie on the filtered posterior, not on the true state path. For unobserved states (e.g., sway velocity in h2), this distribution can deviate substantially. The paper provides no analysis or ablation showing that the transformer's in-context inference is robust to this shift. If the shift degrades the FM's predictions, the entire FM-UKF prediction step is biased, and the reported superiority over CV-UKF and Base-UKF in Table II (h2) may not hold under different sensor configurations or noise levels. This is a correctness risk, not a tuning issue.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces FM-UKF, a state estimator that combines a transformer-based dynamics foundation model with an unscented Kalman filter, using analytically known sensor models instead of learned ones. The authors argue that this hybrid approach retains the generalization of foundation-model dynamics across unseen systems while supporting new sensor configurations without retraining. They contribute an open-source benchmark of 400k container-ship trajectories from 1,000 parameter instances, and evaluate FM-UKF against an end-to-end transformer baseline and three classical UKF variants (Oracle, Base, CV). The main empirical claim, stated in Sec. VI-D, is that in the no-velocity sensor configuration (h2), FM-UKF with a kinematic integrator outperforms the classical baselines, while the end-to-end approach requires retraining per sensor configuration. The paper also reports that the end-to-end baseline, when trained on the same sensor configurations, closely matches the Oracle-UKF.","tokens_in":10913,"tokens_out":4811,"duration_ms":49359,"significance":"If the empirical claims hold, FM-UKF is a meaningful step toward zero-shot state estimation: it uses a learned dynamics prior that generalizes across system instances while leveraging known sensor models, which is a practical architecture for control applications where sensor models are available but system identification is costly. The benchmark itself is a contribution: it is open-sourced, includes a challenging 10-state nonlinear simulation with varied dynamics, and the evaluation uses a held-out split of 100 unseen ships. The paper also makes its code and data available, which supports reproducibility. However, the current evidence for the central claim is incomplete: the headline comparison in Table II lacks statistical uncertainty quantification, the sensor generalization advantage is not tested with a properly zero-shot end-to-end baseline, and a load-bearing assumption about sigma-point context is not validated. These issues need to be addressed before the claim 'FM-UKF outperforms classical baselines' can be considered established.","major_comments":[{"comment":"The central empirical claim that FM-UKF outperforms CV-UKF and Base-UKF in h2 rests on Table II, which reports median MAE and average ranks over 5,000 trajectories. No confidence intervals, error bars, or significance tests are provided, and there are no multiple training seeds. In Fig. 3, the MAE distributions overlap substantially for several states (e.g., phi and psi in h2), so the rank differences in Table II (2.25 for FM-UKF vs 2.75 and 3.38 for Base-UKF and CV-UKF) may not be statistically meaningful. Please provide per-feature paired comparisons, bootstrap confidence intervals across trajectories and/or across the 100 held-out ships, and ideally a repeated-training analysis.","section":"§VI-C/Table II"},{"comment":"The comparison against the end-to-end baseline does not support the paper's central differentiation. Sec. VI-A states that the End2End transformer 'was trained simultaneously on different sensor models, h1 and h2.' The caption to Fig. 3 likewise says the End2End approach matches the oracle 'when trained on identical sensor configurations.' But the paper's claimed advantage is that FM-UKF does not require retraining for new sensor configurations. To demonstrate this, the end-to-end baseline should be trained on h1 only and evaluated on h2, or otherwise the FM-UKF should be evaluated on a sensor configuration not seen during training by either method. The present design conflates dynamics generalization with sensor generalization and weakens the central claim.","section":"§VI-A, Fig. 3"},{"comment":"The load-bearing assumption in Eq. (11) is that 'the sequence of each n-th sigma point trajectory forms a coherent and sufficient context for the transformer to infer the dynamics.' The transformer is trained on true state trajectories (Eq. (12), with small input noise), but at inference it receives sigma-point histories that are filter outputs and can deviate substantially from the true state path, especially for unobserved states such as sway velocity in h2. The paper provides no analysis or ablation quantifying this distribution shift. A direct test would be to feed the UKF's sigma-point histories into the transformer and compare against feeding the true state histories as context, or to evaluate FM-UKF under varying process-noise levels and sensor configurations to see if the learned dynamics become unreliable. Without such evidence, the robustness of the FM prediction step to sigma-","section":"§IV-A, Eq. (11)"},{"comment":"Filter performance depends critically on the process noise covariance Q, which is manually tuned for the classical baselines and then used unchanged for FM-UKF. The authors explicitly state this is a lower bound on FM-UKF performance, but they do not quantify the sensitivity of the ranking to Q. If the relative performance of FM-UKF versus the baselines changes substantially when Q is re-tuned or estimated online, the 'competitive accuracy, effort, and robustness trade-off' claimed in the abstract is not established. A Q-sweep or an adaptive Q method, applied across all filters, would make the comparison fairer and more informative.","section":"§VI-B"}],"minor_comments":[{"comment":"Typo: 'we preserve th KF's interpretability' should be 'the KF's.'","section":"Abstract/Introduction"},{"comment":"Typo: 'not-sufficiently-tuned process nose' should be 'process noise.'","section":"§VI-D"},{"comment":"References [4] and [20] appear to be the same arXiv preprint (Bommasani et al., 'On the Opportunities and Risks of Foundation Models'). Please consolidate or disambiguate.","section":"References"},{"comment":"The description of the MLHP context says Lcontext is 'randomly varied between 64 and 84 during training,' but it is not specified how Lcontext is set at inference time. Since the transformer receives a length-192 trajectory, the masking rule in Eq. (12) must have an inference-time counterpart; please clarify.","section":"§IV-C"},{"comment":"The FM-UKF (no integrator) variant performs drastically worse for positions and yaw, and the authors add a first-principle kinematic integrator to recover performance. This should be discussed more prominently as a design choice: the final method uses a hand-derived kinematic model for position/angle integration, which tempers the 'zero-shot dynamics model' narrative. It would be helpful to state explicitly which states are predicted by the FM and which by the integrator.","section":"§VI-A"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a good fit for the journal and the open-source benchmark is valuable. The core idea is sound, but the empirical validation does not yet support the paper's strongest claims, particularly regarding sensor-configuration generalization and the superiority of FM-UKF over classical baselines. The missing statistical analysis, the unfair end-to-end comparison, and the unvalidated sigma-point distribution shift are all fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a useful engineering paper with a title that overpromises. The real contribution is the clean split between a learned dynamics model and analytically known sensor models inside a UKF—that lets you reuse the dynamics model for new sensor configurations without retraining. That’s a genuine step beyond Busetto et al.’s end-to-end estimator, and the h2 results (no velocity sensors) show FM-UKF beating both a constant-velocity filter and a filter with the correct model structure but detuned parameters. The open benchmark—400k trajectories from 1000 container-ship parameterizations—is also a solid addition.\n\nThe main soft spot is in Section IV-A: the transformer is trained on true state trajectories, but at inference it gets sigma-point histories that have been through UKF corrections. For unmeasured states like sway velocity, those can be far from the truth. The paper states the assumption but never tests it. An ablation with oracle true-state context would settle whether the distribution shift matters. The empirical h2 success suggests the transformer is robust, but as written it’s a correctness risk, not a tuning issue.\n\nSmaller issues: the ‘foundation model’ label is generous for a 1.8M-parameter model trained on one synthetic dataset; and there are no confidence intervals or multiple seeds, which matters because some Table II differences are narrow. The process noise was tuned to the classical baselines, which the authors admit and frame as a lower bound for FM-UKF—that’s honest, but it limits the comparison.\n\nBottom line: I’d send this to peer review, not desk reject, and ask for the oracle-context ablation, error bars, and a more measured title. If you work on learned dynamics for filtering, the benchmark and the hybrid architecture are worth your time.","headline":"Hybrid UKF + learned dynamics is a real step forward, but the sigma-point context shift is untested and the 'zero-shot foundation model' label oversells it.","tokens_in":11344,"tokens_out":4281,"would_cite":true,"duration_ms":42795,"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 single learned dynamics model inside an unscented Kalman filter can estimate the state of unseen container ships zero-shot, beating classical baselines when velocities are not measured.","keywords":["zero-shot state estimation","foundation models","unscented Kalman filter","in-context learning","transformer dynamics model","container ship","system identification","state estimation"],"falsifier":"Run FM-UKF on a ship whose parameters lie at the edge of the training distribution while only position measurements are available, and compare the first 50-step error trajectory against the oracle filter: if the FM-UKF error diverges while the oracle stays bounded, the sigma-point context fails; equivalently, measure the shift between the sigma-point trajectories recorded during filtering and the true-state trajectories in the training set.","tokens_in":10612,"feed_emoji":"🚢","tokens_out":7721,"duration_ms":66118,"temperature":0.7,"pith_summary":"This paper proposes FM-UKF, a state estimator that pairs a pre-trained transformer dynamics model with the unscented Kalman filter (UKF) and its analytically known sensor models. The central claim is that a single learned dynamics model can generalize zero-shot to previously unseen systems—here, 100 container ship parameterizations—while the UKF handles any sensor configuration without retraining. On an open benchmark of container ship models, FM-UKF outperforms a constant-velocity filter and a filter using the correct model structure with detuned parameters, specifically when velocities are not directly measured. The work also shows that an end-to-end transformer baseline, while competitive when trained on the target sensor configuration, must be retrained for each new sensor setup, whereas FM-UKF avoids such retraining by construction. If the claim holds, per-deployment system identification and data collection could be replaced by a single dynamics foundation model combined with a catalog of sensor models.","feed_headline":"Zero-shot filter beats hand-tuned baselines on unseen ships","feed_subtitle":"A transformer learns each ship's motion in context, so new sensor setups need no retraining.","key_machinery":"Sigma-point-as-context prediction inside a UKF. The transform carries sigma points forward through time; instead of a hand-derived dynamics function, a decoder-only transformer (a time-series forecasting architecture) reads the stored sigma-point trajectory plus control inputs and predicts the next state for each sigma point. The UKF's unscented transform propagates uncertainty through this learned predictor, and the analytically known sensor model performs the measurement correction—so the learned part only ever models dynamics, never sensors.","core_discovery":"The central discovery is that the unscented transform's sigma points can double as the context window for a transformer dynamics model. At every filtering step, each sigma-point trajectory is fed through a decoder-only transformer to predict the next state, and the UKF then fuses these predictions with the known sensor model in the correction step. Because the transformer infers the dynamics in context, the same weights serve every ship in the family; because the sensor model is external and analytical, swapping a sensor—say, dropping velocity measurements—requires no new training data. The paper demonstrates this on a benchmark of 1000 container ship parameterizations, reporting that in the","pith_inferences":["The same sigma-point-as-context trick could be transplanted to other ensemble filters (e.g., ensemble Kalman filters or particle filters), where the ensemble naturally provides the context sequence; this might trade some of the UKF's efficiency for robustness to multi-modal posteriors.","Because the authors tuned process noise for the classical baselines and treated the FM-UKF as a lower bound, an obvious extension is to optimize the noise model for the learned predictor—or to have the transformer directly output a covariance—which could close the gap to the oracle and end-to-end methods.","A direct test of the paper's coherence assumption would be to feed the transformer both true-state trajectories and filtered sigma-point trajectories from the same system and compare prediction errors; growing divergence would indicate distribution shift that the current evaluation only indirectly exposes."],"forward_implications":["A single trained dynamics model can replace per-system system identification for a family of systems, since the filtering prediction step is supplied by an in-context learned transformer rather than a hand-derived dynamics function.","New sensor configurations are handled without retraining: the sensor model enters only through the UKF correction, so sensor changes need no new trajectory data.","The successful extension from a two-dimensional toy system to a 10-dimensional nonlinear container ship model suggests the approach scales to realistic nonlinear state spaces.","Releasing the benchmark and dataset gives the community a standardized testbed—1000 parameterizations, 400k trajectories—for zero-shot state estimation research."],"supporting_citations":[{"why":"supplies the end-to-end in-context state estimation approach that FM-UKF extends and the baseline it must beat.","marker":"[10]"},{"why":"supplies the decoder-only time-series forecasting architecture on which the dynamics transformer is based.","marker":"[17]"},{"why":"defines the unscented Kalman filter that carries the prediction and correction steps.","marker":"[2]"},{"why":"supplies the nonlinear container ship dynamics model used to generate the zero-shot benchmark.","marker":"[28]"},{"why":"supplies the 'base set' sigma point generation used inside the UKF.","marker":"[24]"},{"why":"motivates pink-noise control inputs used to generate the training and evaluation trajectories.","marker":"[30]"}],"fun_headline_variants":["Zero-shot ship state estimation with UKF and transformer","FM-UKF: transformer + UKF for zero-shot dynamic systems","Sigma points as context: zero-shot state estimation for new ships","No retraining for new sensors: foundation model UKF filter","Generalizing across ship dynamics with a UKF-transformer hybrid"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"Inside the filter, the transformer receives sigma-point estimates rather than the true state trajectories it was trained on, and the paper does not directly validate that these filtered trajectories stay close enough to the training distribution for the in-context inference to remain sound.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot ship state estimation with UKF and transformer","FM-UKF: transformer + UKF for zero-shot dynamic systems","Sigma points as context: zero-shot state estimation for new ships","No retraining for new sensors: foundation model UKF filter","Generalizing across ship dynamics with a UKF-transformer hybrid"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000155,"raw_usage":{"total_tokens":1023,"prompt_tokens":691,"completion_tokens":332,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":435,"completion_tokens_details":{"reasoning_tokens":246}},"tokens_in":435,"tokens_out":332,"duration_ms":3659,"temperature":1.0,"reasoning_tokens":246,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:18:20.138175+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run FM-UKF on a ship whose parameters lie at the edge of the training distribution while only position measurements are available, and compare the first 50-step error trajectory against the oracle filter: if the FM-UKF error diverges while the oracle stays bounded, the sigma-point context fails; equivalently, measure the shift between the sigma-point trajectories recorded during filtering and the true-state trajectories in the training set.","supporting_citations":[{"cited_title":"In- context learning of state estimators,","cited_arxiv_id":null,"evidence_quote":"supplies the end-to-end in-context state estimation approach that FM-UKF extends and the baseline it must beat."},{"cited_title":"A decoder-only foundation model for time-series forecasting,","cited_arxiv_id":null,"evidence_quote":"supplies the decoder-only time-series forecasting architecture on which the dynamics transformer is based."},{"cited_title":"New extension of the Kalman filter to nonlinear systems,","cited_arxiv_id":null,"evidence_quote":"defines the unscented Kalman filter that carries the prediction and correction steps."},{"cited_title":"On the coupled motion of steering and rolling of a high speed container ship,","cited_arxiv_id":null,"evidence_quote":"supplies the nonlinear container ship dynamics model used to generate the zero-shot benchmark."},{"cited_title":"Bitzer, The UKF exposed: How it works, when it works and when it’s better to sample , Zenodo, 2016","cited_arxiv_id":null,"evidence_quote":"supplies the 'base set' sigma point generation used inside the UKF."},{"cited_title":"Pink noise is all you need: Colored noise exploration in deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"motivates pink-noise control inputs used to generate the training and evaluation trajectories."}],"review_version":1}