pith. sign in

arxiv: 2602.19651 · v2 · submitted 2026-02-23 · 💻 cs.RO · cs.AI· cs.LG

Denoising Particle Filters: Learning State Estimation with Single-Step Objectives

Lennart R\"ostel , Berthold B\"auml This is my paper

Pith reviewed 2026-05-15 20:43 UTC · model grok-4.3

classification 💻 cs.RO cs.AIcs.LG
keywords particle filtersstate estimationdenoising score matchingrobotic systemsBayesian filteringMarkov propertylearning-based methods
0
0 comments X

The pith

Particle filters can be trained on single state transitions by implicitly learning measurement models through denoising score matching.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper introduces a particle filtering method for robotic state estimation that trains its components from individual state transitions rather than full sequences. It learns a denoiser via a score matching objective to capture measurement information implicitly, then uses that denoiser together with a dynamics model to approximate the Bayesian update at each timestep. The approach exploits the Markov property to simplify training while preserving the ability to add prior information or external sensors without retraining the model. In simulation experiments on challenging robotic tasks, it matches the performance of tuned end-to-end sequence models.

Core claim

Measurement models are learned implicitly by minimizing a denoising score matching objective on single transitions; at inference the learned denoiser is combined with a dynamics model to approximately solve the Bayesian filtering equation at each time step, guiding predicted states toward the data manifold informed by measurements.

What carries the argument

The denoising score matching objective, which trains a denoiser on noisy states to implicitly encode measurement likelihoods for guiding particles.

If this is right

  • Training uses only single-step data, eliminating the need to unroll sequences during optimization.
  • The filter remains composable with classical components, allowing prior knowledge and external sensor models to be added at inference time without retraining.
  • Performance on simulated robotic state estimation tasks is competitive with tuned end-to-end sequence models.
  • The Markov property is fully exploited so that each transition can be handled independently during both training and filtering.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The single-step training could make learned filters easier to integrate into existing modular robotic software stacks.
  • Score-based denoising might serve as a drop-in replacement for explicit likelihood models in other recursive estimation settings.
  • Testing on physical hardware would reveal whether accumulated approximation errors remain manageable outside simulation.

Load-bearing premise

That a denoiser trained only on isolated transitions will produce stable and accurate guidance when applied repeatedly across long sequences.

What would settle it

If accuracy on long-horizon robotic trajectories falls significantly below that of an end-to-end trained baseline, or if adding an external sensor model requires retraining the filter.

Figures

Figures reproduced from arXiv: 2602.19651 by Berthold B\"auml, Lennart R\"ostel.

Figure 1
Figure 1. Figure 1: One timestep of inference with Denoising Particle Filters (DnPF). [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: DnPF network structure for efficient denoising inference. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Simulated state estimation tasks used in the experiments. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: DnPF predictions for the Manipulator Spin task. Shown are estimates for the normalized X-component of the object position, ground truth shown in red. At ∼ 2.5s, the manipulator end-effector contacts the object, greatly reducing the space of possible object configurations. After being pulled toward the induced measurement likelihood, DnPF particles follow the (nearly deterministic) dynamics model again, tra… view at source ↗
Figure 5
Figure 5. Figure 5: E. Score Space Sensor Fusion A major advantage of the score-based DnPF formulation is that it allows for the integration of existing sensor models p(ˆyt|xt) without retraining. To see this, consider the joint measurement likelihood p(yt, yˆt|xt) where yˆt denotes the additional external sensor measurements. Assuming condi￾tional independence of yt and yˆt given xt, the score of the joint measurement likeli… view at source ↗
Figure 5
Figure 5. Figure 5: Particle predictions in the Multi-fingered Manipulation task, for one normalized position component (top row) and orientation component (bottom row) respectively. Ground truth is shown in red. The first column shows particles of an open-loop rollout using only the learned dynamics model f. The second column shows a particle rollout of DnPF using only proprioceptive measurements. The particle distribution f… view at source ↗
Figure 6
Figure 6. Figure 6: Analysis of sensor fusion in DnPF on the Multi-fingered Ma [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Performance and inference runtime (GPU) analysis of DnPF with [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
read the original abstract

Learning-based methods commonly treat state estimation in robotics as a sequence modeling problem. While this paradigm can be effective at maximizing end-to-end performance, models are often difficult to interpret and expensive to train, since training requires unrolling sequences of predictions in time. As an alternative to end-to-end trained state estimation, we propose a novel particle filtering algorithm in which models are trained from individual state transitions, fully exploiting the Markov property in robotic systems. In this framework, measurement models are learned implicitly by minimizing a denoising score matching objective. At inference, the learned denoiser is used alongside a (learned) dynamics model to approximately solve the Bayesian filtering equation at each time step, effectively guiding predicted states toward the data manifold informed by measurements. We evaluate the proposed method on challenging robotic state estimation tasks in simulation, demonstrating competitive performance compared to tuned end-to-end trained baselines. Importantly, our method offers the desirable composability of classical filtering algorithms, allowing prior information and external sensor models to be incorporated without retraining.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper proposes a novel particle filtering algorithm for robotic state estimation that trains models on individual state transitions rather than full sequences, exploiting the Markov property. Measurement models are learned implicitly via a denoising score matching objective; at inference, the learned denoiser is combined with a learned dynamics model to approximately solve the Bayesian filtering update at each step by steering particles toward the data manifold. The method is evaluated on challenging simulation tasks and claims competitive performance versus tuned end-to-end baselines while preserving the composability of classical filters for incorporating prior information and external sensors without retraining.

Significance. If the single-step denoising approximation proves stable and accurate over long horizons, the approach would be significant for enabling efficient, interpretable learning-based filtering that avoids sequence unrolling during training and supports modular integration with classical components. The explicit use of the Markov property for single-transition training and the implicit measurement learning via score matching are notable strengths that differentiate it from end-to-end sequence models.

major comments (2)
  1. [Abstract] Abstract: the claim of 'competitive performance' on simulation tasks is unsupported by any quantitative metrics, baseline details, ablation studies, or error analysis, leaving the central assertion that the method approximately solves the Bayesian filtering equation without verifiable experimental grounding.
  2. [Abstract] Abstract (method description): the single-step denoising score matching on isolated transitions is asserted to implicitly recover a measurement model that guides particles to the correct posterior manifold, yet no analysis or bounds are provided on whether the learned score remains consistent with p(z|x) across the predictive distribution or on error accumulation when the approximation is iterated over long sequences in high-dimensional state spaces.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive review and for highlighting the potential significance of our single-step training approach. We address the major comments point by point below, indicating where revisions will be made to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim of 'competitive performance' on simulation tasks is unsupported by any quantitative metrics, baseline details, ablation studies, or error analysis, leaving the central assertion that the method approximately solves the Bayesian filtering equation without verifiable experimental grounding.

    Authors: We agree that the abstract would be strengthened by including specific quantitative support for the performance claim. The full manuscript contains detailed results in Section 5, including RMSE metrics, baseline descriptions, and ablation studies across multiple tasks. In the revision, we will update the abstract to reference key quantitative findings (e.g., average error reductions relative to end-to-end baselines) while preserving brevity. This directly addresses the need for verifiable grounding in the abstract itself. revision: yes

  2. Referee: [Abstract] Abstract (method description): the single-step denoising score matching on isolated transitions is asserted to implicitly recover a measurement model that guides particles to the correct posterior manifold, yet no analysis or bounds are provided on whether the learned score remains consistent with p(z|x) across the predictive distribution or on error accumulation when the approximation is iterated over long sequences in high-dimensional state spaces.

    Authors: This observation correctly identifies a gap in theoretical analysis. The manuscript relies on the Markov property to justify single-step training and demonstrates empirical stability through long-horizon simulations in Section 5, where resampling prevents excessive drift. In the revision, we will expand the abstract and add a short discussion in Section 3 on how the score-matching objective aligns with the measurement model under the predictive distribution. We will also note the absence of formal bounds on high-dimensional error accumulation as a limitation and direction for future work, supported by our practical results. revision: partial

Circularity Check

0 steps flagged

No circularity: derivation relies on standard score matching and particle filter approximation without self-referential reduction

full rationale

The paper's core construction trains a denoiser on isolated state transitions using the standard denoising score matching objective to implicitly represent measurement information, then deploys the learned denoiser together with a separately learned dynamics model inside a particle filter to approximate the Bayesian update at each step. No equation or claim reduces the claimed performance or the implicit measurement model to a quantity defined by the same fitted objective; the single-step training exploits the Markov property explicitly stated as an assumption, and the inference procedure is presented as an approximation whose validity is evaluated empirically rather than guaranteed by construction. No self-citation chains or imported uniqueness theorems appear in the provided abstract or description to bear the central load. The method is therefore self-contained against external benchmarks of score matching and classical filtering.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that robotic systems obey the Markov property sufficiently for single-step training to suffice, plus the implicit assumption that score matching yields a usable measurement model. No free parameters or invented entities are introduced in the abstract description.

axioms (1)
  • domain assumption Robotic systems satisfy the Markov property, so single state transitions contain all necessary information for training.
    Explicitly invoked in the abstract as 'fully exploiting the Markov property in robotic systems'.

pith-pipeline@v0.9.0 · 5477 in / 1328 out tokens · 26235 ms · 2026-05-15T20:43:15.329420+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

37 extracted references · 37 canonical work pages · 4 internal anchors

  1. [1]

    General in-hand object rotation with vision and touch,

    H. Qiet al., “General in-hand object rotation with vision and touch,” arXiv [cs.RO], Sep. 2023

    work page 2023

  2. [2]

    Learning a state estimator for tactile in-hand manipulation,

    L. R ¨ostel, L. Sievers, J. Pitz, and B. B¨auml, “Learning a state estimator for tactile in-hand manipulation,” in2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Oct. 2022

    work page 2022

  3. [3]

    Particle filter networks with appli- cation to visual localization,

    P. Karkus, D. Hsu, and W. S. Lee, “Particle filter networks with appli- cation to visual localization,” inProceedings of The 2nd Conference on Robot Learning. PMLR, 2018

    work page 2018

  4. [4]

    MegaParticles: Range-based 6-DoF monte carlo localization with GPU-accelerated stein particle filter,

    K. Koide, S. Oishi, M. Yokozuka, and A. Banno, “MegaParticles: Range-based 6-DoF monte carlo localization with GPU-accelerated stein particle filter,” in2024 IEEE International Conference on Robotics and Automation (ICRA), vol. 29. IEEE, May 2024

    work page 2024

  5. [5]

    Probabilistic robotics,

    S. Thrun, “Probabilistic robotics,”Communications of the ACM, 2002

    work page 2002

  6. [6]

    A tutorial on particle filtering and smoothing: Fifteen years later,

    A. Doucet and A. M. Johansen, “A tutorial on particle filtering and smoothing: Fifteen years later,”Handbook of Nonlinear Filtering, 2008

    work page 2008

  7. [7]

    Backprop KF: Learning discriminative deterministic state estimators,

    T. Haarnoja, A. Ajay, S. Levine, and P. Abbeel, “Backprop KF: Learning discriminative deterministic state estimators,”arXiv [cs.LG], May 2016

    work page 2016

  8. [8]

    Differentiable particle filters: End-to-end learning with algorithmic priors,

    R. Jonschkowski, D. Rastogi, and O. Brock, “Differentiable particle filters: End-to-end learning with algorithmic priors,”arXiv [cs.LG], May 2018

    work page 2018

  9. [9]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,”Neural computation, vol. 9, no. 8, 1997

    work page 1997

  10. [10]

    Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation

    K. Choet al., “Learning phrase representations using rnn encoder-decoder for statistical machine translation,”arXiv preprint arXiv:1406.1078, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  11. [11]

    Attention is all you need,

    A. Vaswaniet al., “Attention is all you need,”Advances in neural information processing systems, vol. 30, 2017

    work page 2017

  12. [12]

    How to train your differentiable filter,

    A. Kloss, G. Martius, and J. Bohg, “How to train your differentiable filter,”Autonomous Robots, 2021

    work page 2021

  13. [13]

    Learning dynamics models for model predictive agents,

    M. Lutteret al., “Learning dynamics models for model predictive agents,”arXiv [cs.LG], Sep. 2021

    work page 2021

  14. [14]

    The unscented particle filter,

    R. van der Merwe, A. Doucet, N. de Freitas, and E. Wan, “The unscented particle filter,”Advances in Neural Information Processing Systems, vol. 13, 2000

    work page 2000

  15. [15]

    The manifold particle filter for state estimation on high-dimensional implicit manifolds,

    M. C. Kovalet al., “The manifold particle filter for state estimation on high-dimensional implicit manifolds,” in2017 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2017

    work page 2017

  16. [16]

    State estimation in contact-rich manipulation,

    F. Wirnshoferet al., “State estimation in contact-rich manipulation,” in 2019 International Conference on Robotics and Automation (ICRA). IEEE, May 2019

    work page 2019

  17. [17]

    Kernel embedded nonlinear observational mappings in the variational mapping parti- cle filter,

    M. Pulido, P. J. vanLeeuwen, and D. J. Posselt, “Kernel embedded nonlinear observational mappings in the variational mapping parti- cle filter,” inInternational Conference on Computational Science. Springer, 2019

    work page 2019

  18. [18]

    Stein particle filter for nonlinear, non-gaussian state estimation,

    F. A. Maken, F. Ramos, and L. Ott, “Stein particle filter for nonlinear, non-gaussian state estimation,”IEEE Robotics and Automation Letters, 2022

    work page 2022

  19. [19]

    Stein variational gradient descent: A general purpose bayesian inference algorithm,

    Q. Liu and D. Wang, “Stein variational gradient descent: A general purpose bayesian inference algorithm,”arXiv [stat.ML], Aug. 2016

    work page 2016

  20. [20]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,”Advances in neural information processing systems, vol. 33, 2020

    work page 2020

  21. [21]

    Score-based generative modeling through stochastic differential equations,

    Y . Songet al., “Score-based generative modeling through stochastic differential equations,”arXiv [cs.LG], Nov. 2020

    work page 2020

  22. [22]

    Diffusion models beat GANs on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat GANs on image synthesis,”arXiv [cs.LG], 2021

    work page 2021

  23. [23]

    Classifier-Free Diffusion Guidance

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,”arXiv preprint arXiv:2207.12598, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  24. [24]

    Diffusion policy: Visuomotor policy learning via action diffusion,

    C. Chiet al., “Diffusion policy: Visuomotor policy learning via action diffusion,” inProceedings of Robotics: Science and Systems (RSS), 2023

    work page 2023

  25. [25]

    Fighting uncertainty with gradients: Offline rein- forcement learning via diffusion score matching,

    H. T. Suhet al., “Fighting uncertainty with gradients: Offline rein- forcement learning via diffusion score matching,” inConference on Robot Learning. PMLR, 2023

    work page 2023

  26. [26]

    Score-based data assimilation,

    F. Rozet and G. Louppe, “Score-based data assimilation,”Neural Inf Process Syst, Jun. 2023

    work page 2023

  27. [27]

    A new approach to linear filtering and prediction problems,

    R. E. Kalman, “A new approach to linear filtering and prediction problems,”Journal of Basic Engineering, vol. 82, no. 1, pp. 35–45, 1960

    work page 1960

  28. [28]

    Deep unsupervised learning using nonequilibrium thermodynamics,

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermodynamics,” inInternational conference on machine learning. pmlr, 2015

    work page 2015

  29. [29]

    Generative modeling by estimating gradients of the data distribution,

    Y . Song and S. Ermon, “Generative modeling by estimating gradients of the data distribution,” 2019

    work page 2019

  30. [30]

    Flow Matching for Generative Modeling

    Y . Lipmanet al., “Flow matching for generative modeling,”arXiv preprint arXiv:2210.02747, Oct. 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  31. [31]

    An introduction to flow matching and diffusion models

    P. Holderrieth and E. Erives, “An introduction to flow matching and diffusion models,”arXiv preprint arXiv:2506.02070, Jun. 2025

    work page arXiv 2025

  32. [32]

    Toward practical N2 monte carlo: The marginal particle filter,

    M. Klaas, N. de Freitas, and A. Doucet, “Toward practical N2 monte carlo: The marginal particle filter,”arXiv [stat.CO], Jul. 2012

    work page 2012

  33. [33]

    Film: Visual reasoning with a general conditioning layer,

    E. Perezet al., “Film: Visual reasoning with a general conditioning layer,” inProceedings of the AAAI conference on artificial intelligence, vol. 32, no. 1, 2018

    work page 2018

  34. [34]

    Layer Normalization

    J. L. Ba, J. R. Kiros, and G. E. Hinton, “Layer normalization,”arXiv preprint arXiv:1607.06450, 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  35. [35]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,”arXiv [cs.LG], Oct. 2020

    work page 2020

  36. [36]

    On the continuity of rotation representations in neural networks,

    Y . Zhouet al., “On the continuity of rotation representations in neural networks,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019

    work page 2019

  37. [37]

    Mujoco: A physics engine for model-based control,

    E. Todorov, T. Erez, and Y . Tassa, “Mujoco: A physics engine for model-based control,” in2012 IEEE/RSJ international conference on intelligent robots and systems. IEEE, 2012

    work page 2012