Differentiable Bayesian Filtering for Structured State Estimation

31 concepts27 papers1 proposal

State estimation—the challenge of tracking an agent's hidden state from a sequence of noisy sensor measurements—is fundamental to robotics, tracking, and control. Traditionally, this problem is solved using classical Bayesian filters, such as the Kalman Filter or Particle Filter. These filters rely on explicit, hand-crafted mathematical models of how a system evolves over time (the dynamics model) and how sensors measure the state (the observation model). While highly interpretable and excellent at tracking uncertainty, classical filters struggle when dealing with raw, high-dimensional inputs like camera pixels, raw audio, or complex tactile feedback.

Conversely, deep neural networks excel at processing raw data but lack structural priors. When generic recurrent networks (like LSTMs) are tasked with state estimation, they treat it as a black-box mapping. This often leads to overfitting, poor generalization, and an inability to explicitly represent uncertainty.

Differentiable state estimation bridges this gap. Instead of replacing classical filters with neural networks, it embeds the mathematical steps of classical filters directly into a neural network's computational graph. By making the entire filtering loop differentiable, we can use backpropagation to optimize neural network feature extractors, transition dynamics, and noise parameters end-to-end based on the final tracking performance. As Alina Kloss demonstrates in How to Train Your Differentiable Filter, leveraging these recursive probabilistic priors provides a strong inductive bias that consistently outperforms unstructured recurrent networks.


End-to-End Optimization vs. Decoupled Modeling

To understand why differentiable filtering works, consider how traditional systems are tuned. Typically, one learns or identifies a transition model $p(x_t \mid x_{t-1})$ and an observation model $p(z_t \mid x_t)$ independently—for instance, by maximizing the likelihood of isolated sensor readings. However, small errors in these decoupled generative models compound over long temporal sequences, leading to catastrophic tracking failures.

By formulating the filter as a sequence of differentiable operations, the system optimizes for filtering performance directly. As Arun Venkatraman et al. show in Inference Machines for Nonparametric Filter Learning, this discriminative approach allows the filter to ignore modeling errors that do not impact the final state estimate, bypassing the need to perfectly reconstruct complex environments. Gradients flow backwards from the final state estimation error through the update and prediction steps, directly shaping how the neural network encodes raw observations.

Raw Observations (z_t) ---> [ Neural Feature Encoder ] 
                                      |
                                      v
Previous Belief (b_{t-1}) ---> [ Differentiable Filter Step ] ---> Updated Belief (b_t)
                                      |
                                      v
                           Loss = Distance(b_t, Ground Truth)
                                      |
       (Gradients Backpropagated Through Time to Optimize Encoder & Dynamics)

Differentiable Kalman Filtering: Learning Noise and Dynamics

In a standard Kalman Filter, tracking accuracy depends heavily on transition matrices, observation matrices, and noise covariance matrices ($Q$ and $R$). Manually tuning these parameters for complex sensors is notoriously difficult.

In a Differentiable Kalman Filter (DKF), these matrices are parameterized by neural networks. For instance, Huseyin Coskun et al. in Long Short-Term Memory Kalman Filters: Recurrent Neural Estimators for Pose Regularization use LSTMs to output the regularizing parameters of a Kalman Filter, significantly improving tracking under complex sensor noise. Similarly, Wei Xue et al. apply this to speech enhancement in Neural Kalman Filtering for Speech Enhancement, replacing rigid linear assumptions with non-linear neural components while preserving the filter's structural expert knowledge.

This architecture also allows the system to learn sensor reliability automatically. In Learning Interpretable BEV Based VIO without Deep Neural Networks, Zexi Chen et al. backpropagate errors through an Unscented Kalman Filter to optimize noise covariance parameters directly, resulting in an interpretable system with only four trainable parameters.

When scaled to distributed networks or multi-sensor systems, gradients can flow across nodes. Colin Samplawski et al. demonstrate in Heteroskedastic Geospatial Tracking with Distributed Camera Networks that tracking errors from a central fusion node can be backpropagated to individual camera adapters, optimizing local feature extraction based on global tracking performance. Similarly, Chien-Yi Wang et al. use a differentiable multi-sensor Kalman Filter in Probabilistic 3D Multi-Object Cooperative Tracking for Autonomous Driving via Differentiable Multi-Sensor Kalman Filter to learn optimal weightings for multi-source data sharing.


Differentiable Particle Filtering: Overcoming the Resampling Discontinuity

While Kalman filters work well for unimodal distributions, non-linear and multi-modal tracking problems require Particle Filters (PFs). A particle filter represents the belief distribution using a set of discrete samples (particles). However, particle filters introduce a major mathematical challenge for backpropagation: the resampling step.

Resampling selects particles stochastically based on their weights, which is a discrete, non-differentiable operation. Without a differentiable workaround, gradients cannot flow back through time across the resampling step.

[ Particles & Weights ] ---> [ Discrete Resampling (Non-Differentiable) ] ---> Gradient Blocked ❌

[ Particles & Weights ] ---> [ Entropy-Regularized Optimal Transport ]   ---> Gradient Flows  ===

To solve this, researchers use smooth approximations. In Differentiable Particle Filtering via Entropy-Regularized Optimal Transport, Adrien Corenflos et al. replace discrete resampling with a smooth transport mechanism based on entropy-regularized Optimal Transport (the Sinkhorn algorithm). This allows the entire filtering loop to remain fully differentiable.

By making particle filters differentiable, neural networks can learn complex observation models directly from raw data within a multi-modal tracking framework. For example, David Hsu et al. introduce PF-net in Particle Filter Networks with Application to Visual Localization, embedding a particle filter into a recurrent neural network to map camera images to 2D floor maps. Similarly, Peter Karkus et al. leverage this in Discriminative Particle Filter Reinforcement Learning for Complex Partial observations to optimize a tracking filter jointly with a reinforcement learning policy.


Expanding the Scope: Factor Graphs and Belief-Space Planning

The paradigm extends beyond standard recursive filtering to other state estimation structures, such as factor graphs and belief-space planning.

In Learning Dynamics of a Ball with Differentiable Factor Graph and Roto-Translational Invariant Representations, Qingyu Xiao et al. integrate a differentiable factor graph into the training loop, allowing the system to refine noisy initial observations to better align with learned physical laws.

Furthermore, maintaining a structured, differentiable belief distribution enables agents to plan actions that actively reduce uncertainty. In QMDP-Net: Deep Learning for Planning under Partial Observability, Peter Karkus et al. embed a differentiable POMDP solver into a neural network, allowing the system to learn internal representations for state estimation and planning simultaneously. This active uncertainty reduction is also demonstrated by JEAN-FRANÇOIS TREMBLAY et al. in Learning active tactile perception through belief-space control, where a differentiable filtering layer helps a robot plan exploratory actions to identify unknown object properties.

Key Foundational Work

Backprop KF: Learning Discriminative Deterministic State Estimators (2016) — contributed 2 concept elements to this direction.
Inference Machines for Nonparametric Filter Learning (2016) — contributed 2 concept elements to this direction.
QMDP-Net: Deep Learning for Planning under Partial Observability (2017) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_differentiable_state_estimation_495f7937,
  title = {differentiable state estimation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/differentiable-state-estimation/rdr_495f793786a64a86ba4f9d5cba3631f0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_differentiable_state_estimation_495f7937_v1,
  title = {differentiable state estimation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/differentiable-state-estimation/rdr_495f793786a64a86ba4f9d5cba3631f0/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake