Path Integral Particle Filtering for Hybrid Systems via Saltation Matrices

Karthik Shaji, Sreeranj Jayadevan, Bo Yuan, Hongzhe Yu, Haomin ZhouarXiv 2026 · DOI · arXiv

TL;DR

Estimating the state of robots that bounce, collide, or switch modes (hybrid systems) is notoriously difficult because traditional filters struggle with discontinuous "jumps" in dynamics. This paper introduces the Salted Path Integral Particle Filter (SPIPF), which combines the mathematical duality of control and estimation with "Saltation Matrices" to accurately track uncertainty through impacts, outperforming standard filters even with fewer particles.

What Problem Does This Paper Address?

In the world of robotics and aerospace, many systems are "hybrid"—they exhibit continuous motion (like a robot flying through the air) interrupted by discrete events (like the robot's foot hitting the ground). These events, or "mode transitions," create massive headaches for state estimation:

  1. Discontinuity: Standard filters like the Extended Kalman Filter (EKF) rely on smooth math (Jacobians). When a collision occurs, the math becomes non-smooth or undefined.
  2. Non-Gaussian Noise: Impacts often introduce "outlier" measurements or noise distributions that aren't a neat bell curve, causing Gaussian-based filters to fail.
  3. Weight Degeneracy: Traditional particle filters often see their "particles" (hypothesized states) die off because they can't reconcile the sudden change in dynamics with the sensor data, leading to a "weight degeneracy" problem where only one or two particles carry all the mathematical weight.

Methods

The authors propose a novel framework called SPIPF that bridges two advanced concepts:

  • Path Integral Duality: Instead of treating estimation as a simple Bayesian update, they frame it as an optimal control problem. By finding a "control law" that guides particles toward the most likely trajectory, the filter becomes much more robust to outliers and non-Gaussian noise.
  • Saltation Matrices: This is the core innovation for handling impacts. A Saltation Matrix is a mathematical tool used to linearize dynamics across a jump. Unlike a standard Jacobian, it accounts for the fact that different particles might hit a boundary (like the floor) at slightly different times. This allows the filter to map uncertainty through a collision accurately.
  • Salted iLQR: They use a modified version of the Iterative Linear Quadratic Regulator (iLQR) to solve the underlying control problem for each particle, incorporating the Saltation Matrices during the "backward pass" of the algorithm.
  • Sliding Window: To keep the algorithm computationally efficient, they use a sliding window approach, only looking back at a specific chunk of the trajectory rather than the entire history.

Datasets

The researchers evaluated their algorithm on two simulated benchmarks representing classic hybrid systems:

  1. 1D Bouncing Ball: A linear system with a discrete impact event (hitting the ground) and a coefficient of restitution.
  2. Spring Loaded Inverted Pendulum (SLIP): A highly nonlinear and unstable model used to simulate legged locomotion (bipedal/quadrupedal robots). This system features complex state transitions and non-identity reset maps.

The performance was compared against a standard Sequential Importance Resampling (SIR) multi-mode particle filter and an ablation version of their own algorithm (SPIPF-0) that lacks the optimal control guidance.

Key Insights and Ideas

  • Control as Estimation: By treating the "best" state estimate as the solution to a control problem, the filter can "look ahead" and "look back" within its sliding window. This makes it significantly harder for a single bad sensor reading to throw the filter off.
  • Handling the "Jump": The Saltation Matrix is superior to simply differentiating a reset map because it captures the sensitivity of the transition time. This prevents the filter from "losing" particles during the exact moment of impact.
  • Efficiency through Guidance: Because the particles are "guided" by an optimal controller, the filter requires far fewer particles to achieve high accuracy. The authors found that SPIPF with just 10 particles often outperformed standard filters using 50 particles.

Key Results

  • Accuracy: SPIPF consistently achieved lower Mean Squared Error (MSE) and lower MSE Covariance compared to SIR and SPIPF-0 across both the Bouncing Ball and SLIP systems.
  • Robustness: The algorithm maintained a much higher Effective Sample Size (ESS). In the SLIP system, standard filters saw their ESS crash to near zero after a transition (indicating degeneracy), while SPIPF remained stable.
  • Mode Prediction: SPIPF showed higher accuracy in predicting which "mode" the system was in (e.g., flight vs. stance) compared to the baselines.

Limitations and Future Work

  • Sensitivity to Guard Detection: The authors noted that the filter is sensitive to "guard detection"—the logic that determines exactly when a transition occurs. If the sliding window is too large in a highly unstable system, it can trigger transitions prematurely.
  • Computational Overhead: While more efficient in terms of particle count, running an iLQR optimization for every particle is still computationally heavier than a simple Kalman update.
  • Future Directions: The authors suggest exploring multi-threading to run particle rollouts in parallel, implementing adaptive resampling, and eventually testing the framework on physical hardware (like legged robots) combined with learning-based system identification.

Summary updated

Abstract

We present an optimal-control-based particle filtering method for state estimation in hybrid systems that undergo intermittent contact with their environments. We follow the path integral filtering framework that exploits the duality between the smoothing problem and optimal control. We leverage saltation matrices to map out the uncertainty propagation during contact events for hybrid systems. The resulting path integral optimal control problem allows for a state estimation algorithm robust to outlier effects, flexible to non-Gaussian noise distributions, that also handles the challenging contact dynamics in hybrid systems. This work offers a computationally efficient and reliable estimation algorithm for hybrid systems with stochastic dynamics. We also present extensive experimental results demonstrating that our approach consistently outperforms strong baselines across multiple settings.
BibTeX
@misc{shaji2026path,
  author = {Karthik Shaji and Sreeranj Jayadevan and Bo Yuan and Hongzhe Yu and Haomin Zhou},
  title = {{Path Integral Particle Filtering for Hybrid Systems via Saltation Matrices}},
  year = {2026},
  eprint = {2603.01176},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.01176},
}
Figure