In classical robotics and control, state estimation is handled by "white-box" algorithms like the Kalman Filter or Particle Filter. These models are mathematically rigorous and interpretable, but they struggle with high-dimensional, "messy" data like raw pixels or complex sensor noise. Conversely, deep learning excels at feature extraction but often acts as a "black box," ignoring the physical constraints and recursive logic that govern moving systems.
A growing body of research seeks to bridge this gap by implementing classical filtering algorithms as differentiable layers within neural networks. By making the internal steps of a filter—such as the Bayesian update or particle resampling—mathematically differentiable, researchers can train entire perception-to-estimation pipelines end-to-end.
The Filter as a Neural Layer
The core insight of this research direction is that the equations governing state estimation are essentially a sequence of matrix operations. In Backprop KF, Tuomas Haarnoja and Sergey Levine demonstrated that by reframing the Kalman Filter as a differentiable computation graph, a model can learn to extract low-dimensional features from images that are specifically optimized for the filter’s mathematical requirements.
Instead of manually tuning noise covariance matrices ($Q$ and $R$), which is notoriously difficult, these parameters are treated as learnable weights. Ido Greenberg and Netanel Yannay showed in Optimization or Architecture: How to Hack Kalman Filtering that simply optimizing a standard Kalman Filter via backpropagation often allows it to outperform complex LSTMs, because the filter provides a superior "inductive bias" for temporal tasks.
Overcoming the Differentiability Hurdle
While the Kalman Filter is naturally composed of differentiable linear algebra, the Particle Filter presents a major challenge: the resampling step. Traditionally, resampling involves a discrete, stochastic selection of particles based on their weights, which creates a "dead end" for gradients during backpropagation.
To solve this, researchers have introduced several "smooth" alternatives to discrete resampling:
- Optimal Transport: Adrien Corenflos and Arnaud Doucet utilized entropy-regularized Optimal Transport (the Sinkhorn algorithm) to transform particles in a way that is fully differentiable, as detailed in Differentiable Particle Filtering via Entropy-Regularized Optimal Transport.
- Learnable Resamplers: Kevin Murphy and Rico Jonschkowski proposed the Particle Transformer, which uses a neural network to learn the resampling strategy itself.
- Diffusion Models: More recently, Zheng Zhao introduced diffusion differentiable resampling, treating the resampling process as a continuous stochastic flow that gradients can easily pass through.
Why It Works: Structured Inductive Bias
The primary advantage of these hybrid models is that they enforce the Markov assumption—the idea that the future state depends only on the current state and an action. Standard Recurrent Neural Networks (RNNs) like LSTMs are theoretically capable of learning this, but they often overfit or fail to maintain a consistent physical state over long periods.
By embedding a filter, the architecture forces the network to maintain an explicit "belief" about the world. In QMDP-Net, Peter Karkus and David Hsu showed that embedding a differentiable POMDP solver allows an agent to navigate uncertain environments more effectively than generic networks because the model "knows" it must update its belief based on new observations.
Similarly, DynaNet by Changhao Chen and colleagues uses a "Neural Kalman" framework to ensure motion estimation remains physically interpretable while still benefiting from the feature-learning power of deep LSTMs.
Applications in Complex Environments
Differentiable filters are particularly effective in robotics and high-precision tracking where sensor data is multimodal and noisy.
- Multimodal Fusion: Michelle A. Lee and Jeannette Bohg demonstrated in Multimodal Sensor Fusion with Differentiable Filters that these models can learn to weigh vision, touch, and proprioception end-to-end, outperforming modularly designed filters.
- Cooperative Tracking: In DMSTrack, Chien-Yi Wang used a differentiable multi-sensor Kalman Filter to allow autonomous vehicles to share learned uncertainty estimates, significantly improving collective tracking accuracy.
- Scientific Modeling: Beyond robotics, Francisco Barreras and Victor M. Preciado applied these concepts to epidemic forecasting in AutoEKF, using automatic differentiation to calibrate high-dimensional stochastic models of disease spread.
By treating classical algorithms not as rigid formulas but as differentiable templates, this research direction enables a new class of "physics-aware" AI that combines the interpretability of the past with the learning capacity of the present.