Self-Supervised Hybrid Kalman Filter Boosts Tracking and Classification
New filter adapts without large datasets—works with few or many measurements.
Kalman filters are foundational for tracking and control but break down under model mismatch or poor noise tuning. A new arXiv paper from Lee, Ahmed, and Russell introduces a self-supervised Hybrid Adaptive Kalman Filter (HAKF) that learns structured corrections to both the dynamics model and the process noise covariance directly from measurement data—no ground-truth labels or large datasets required.
By preserving the filter's probabilistic structure, HAKF maintains consistent uncertainty estimates and allows computation of the innovation likelihood. This likelihood is then used for generalized Bayesian inference to classify the underlying system model (e.g., which dynamics regime is active). Experiments on real-world robotics datasets and simulations show improved estimation accuracy and statistical consistency compared to standard Kalman filters and prior adaptive methods. The method excels in low-data scenarios yet scales well with more data, making it practical for autonomous systems where labeled data is scarce or conditions change rapidly.
- Self-supervised learning of dynamics and noise corrections from measurements alone, eliminating need for large labeled datasets.
- Preserves probabilistic consistency, enabling innovation likelihood computation for Bayesian model classification.
- Demonstrated robust performance on real-world and simulated data across both low-data and large-data regimes.
Why It Matters
Enables more reliable tracking and classification in autonomous systems without costly data labeling.