💻 computer science

A Sliding-Window Filter for Online Continuous-Time Continuum Robot State Estimation

This paper presents the first stochastic sliding-window filter for continuum robots that enables accurate, online, continuous-time state estimation while operating at faster-than-real-time speeds, overcoming the limitations of existing discrete-time approximations and offline methods.

Original authors: Spencer Teetaert, Sven Lilge, Jessica Burgner-Kahrs, Timothy D. Barfoot

Published 2026-05-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Spencer Teetaert, Sven Lilge, Jessica Burgner-Kahrs, Timothy D. Barfoot

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a continuum robot as a super-flexible, snake-like machine. Unlike a standard robot arm made of stiff metal joints, this robot can bend, twist, and curl into almost any shape, allowing it to squeeze into tight, dangerous, or delicate spaces—like inside a human body for surgery or through the rubble of a collapsed building.

The big problem with these robots is that they are hard to track. Because they bend so much, it's difficult to know exactly where every part of the snake is at any given moment. If you just guess based on how you moved the controls (open-loop), you'll be wrong because the robot might bump into something or the material might stretch slightly. To fix this, we need a "state estimator"—a smart computer program that looks at sensor data and guesses the robot's true shape.

The Three Ways to Guess the Shape

The paper compares three different ways to make these guesses:

The "Instant" Filter (The Old Way):
Imagine you are driving a car and only looking at the road right in front of your bumper. You make a decision, move forward, and then look again. This is fast and happens in real-time, but if you take a sharp turn, you might miss the curve because you didn't look ahead or remember the turn you just made. In robotics, this is called a filter. It's fast but often a bit inaccurate because it forgets the recent past.
The "Perfect" Batch Solver (The Offline Way):
Imagine you drive the whole route, stop the car, and then look at a video recording of the entire trip from start to finish. You can now see exactly where you were at every second and correct your mistakes. This is incredibly accurate. In robotics, this is called batch optimization. However, you can't do this while driving; you have to wait until the job is done. It's too slow to use while the robot is moving.
The "Sliding Window" (The New Solution):
This is what the authors created. Imagine you are driving, but instead of just looking at the bumper, you have a rearview mirror that shows the last 10 seconds of your trip.
- You look at where you are now.
- You also look at where you were 1, 2, and 3 seconds ago.
- You use all that recent history to make a much smarter guess about where you are right now.
- As you move forward, the "window" slides: the oldest second drops off the back, and a new second is added to the front.

This method, called a Sliding-Window Filter (SWF), tries to get the best of both worlds: the speed of the "Instant" filter and the accuracy of the "Perfect" batch solver.

What the Paper Actually Did

The researchers built a computer program that acts like this sliding rearview mirror for a flexible robot. Here is what they found:

It's Faster Than Real-Time: The computer can calculate the robot's shape faster than the robot is actually moving. This means the robot can use this information immediately to control itself without getting "behind."
It's More Accurate: By looking at a tiny bit of history (about 0.1 seconds, or a few frames of a video), the robot's guess about its shape became much more accurate than the old "Instant" method.
It's Almost as Good as the "Perfect" Method: While the "Perfect" batch method (looking at the whole history) is slightly better, the new sliding window method gets very close to that level of accuracy while still being fast enough to use while the robot is working.
It Handles Mistakes: They tested the robot with sensors that sometimes lose signal (like a GPS dropping out in a tunnel). The sliding window method was able to keep tracking the robot's shape using other sensors (like gyroscopes) until the main signal came back, whereas simpler methods struggled more.

The Trade-Off

The only "cost" of this new method is a tiny bit of latency (a slight delay). Because the system looks at a small window of the past to make its best guess, the "current" position it reports is actually from a fraction of a second ago. However, the authors found this delay is so small (less than a tenth of a second) that it doesn't matter for most practical uses.

Summary

In short, the paper introduces a new "smart rearview mirror" for flexible robots. It allows the robot to know exactly what shape it is in, in real-time, with high accuracy, by briefly looking at its own recent history. This makes it possible to control these tricky, snake-like robots more safely and precisely in the real world.

Technical Summary: A Sliding-Window Filter for Online Continuous-Time Continuum Robot State Estimation

Problem Statement
Continuum robots (CRs) are flexible manipulators capable of navigating confined spaces, yet their control relies on accurate state estimation. Existing probabilistic methods face a dichotomy:

Filtering methods (e.g., Extended Kalman Filters) are computationally efficient and suitable for online operation but often suffer from reduced accuracy due to linearization approximations that fail to fully capture the Markov structure of nonlinear systems.
Batch smoothing methods offer superior accuracy by jointly optimizing all states but are typically restricted to offline operation, as they require retrospective access to all measurements and do not scale well for real-time control.
Continuous-time approaches have shown promise for CRs but have previously been limited to offline batch optimization. Furthermore, existing sliding-window attempts for CRs have relied on simplified, non-probabilistic models (e.g., constant-curvature or rigid-link approximations) or lacked stochastic representations to quantify uncertainty.

Methodology
The authors propose a Sliding-Window Filter (SWF) designed specifically for the online, continuous-time state estimation of CRs. The method is built upon a factor-graph framework derived from recent batch optimization techniques.

Estimation Framework: The system estimates the robot's pose $T(s,t)$ $T (s, t)$ , velocity $\varpi(s,t)$ $ϖ (s, t)$ , and strain $\epsilon(s,t)$ $ϵ (s, t)$ continuously over arc length $s$ $s$ and time $t$ $t$ . The problem is formulated as a Maximum A Posteriori (MAP) estimation, minimizing the negative log-likelihood of prior and measurement factors.
- Priors: Utilize an approximate Cosserat rod model based on a "white-noise-on-acceleration" motion prior.
- Measurements: Incorporate tip pose measurements and gyroscope data.
Sliding-Window Formulation: The filter maintains a fixed-length time window containing states $x_{a:k}$ $x_{a : k}$ .
- Factorization: The joint posterior is decomposed into marginalized prior factors (encapsulating history before the window), spatial factors (along the robot's arc length), and temporal motion/measurement factors (within the window).
- Marginalization: To maintain a fixed window size while propagating information, the oldest time step is marginalized out using the Schur complement in information form. This creates an effective prior for the new window, preserving the Markov property while mitigating errors inherent in strict Markov assumptions.
Implementation Details:
- Window Expansion: New states are initialized by averaging a spatial prior-consistent estimate and a temporal prior-consistent estimate.
- State Extraction: The mean and covariance are extracted from the back of the window to utilize all available measurements, introducing a small latency (typically $\leq 0.1$ s).
- Covariance Handling: To address discontinuities caused by re-linearization, joint covariance estimates between neighboring time steps are stored as states are locked.

Key Contributions

First Probabilistic SWF for CRs: This work presents the first stochastic sliding-window filter specifically designed for continuum robots, enabling the estimation of variable-curvature shapes over both space and time.
Online Continuous-Time Operation: The method enables continuous-time state estimation to run online, overcoming the offline restriction of previous batch continuous-time methods.
Balanced Performance: It achieves a compromise between the computational efficiency of filtering and the accuracy of batch smoothing, operating at faster-than-real-time speeds while maintaining stochastic uncertainty representations.
Open-Source Availability: An open-source implementation is provided to the community to facilitate further research.

Results
The method was validated on a real 3D-printed tendon-driven continuum robot across five distinct trajectories (including free-space motion, fast contact, and impulse interactions).

Accuracy: The SWF (with a 0.1s window) significantly outperformed the filter baseline (0s window) in terms of Tip Position and Rotation Root Mean Square Error (RMSE). For example, in the "Fast Contact" trajectory, the SWF reduced position RMSE from 1.75% to 1.29%.
Comparison to Batch: The SWF achieved accuracy comparable to the full 10s batch optimization baseline (e.g., 1.29% vs. 1.29% RMSE in Fast Contact) but with drastically reduced computational cost.
Computational Efficiency: The SWF operated at approximately 10ms per time step (100Hz), well within real-time constraints (33.3ms threshold), whereas the batch method required ~1.3 seconds per step.
Window Size Sensitivity: Results indicated diminishing returns in accuracy for window sizes larger than 0.1s. Notably, even very small windows (0.033s) provided significant accuracy improvements over the standard filter.
Robustness: The filter demonstrated the ability to track the robot during fast motions and recover quickly from temporary sensor dropouts, with uncertainty estimates correctly expanding during data loss.

Significance and Claims
The authors claim this work represents a "promising direction" for the field by bridging the gap between filtering and smoothing. The primary significance lies in enabling online, continuous-time, stochastic state estimation for continuum robots, a capability previously unavailable.

The paper modestly notes that while the method introduces a small latency and slightly reduced smoothness compared to full batch methods, it maintains the continuous-time properties of the original framework. The authors suggest that the remaining computational headroom (running at >200Hz equivalent for the filter-only variant) could support downstream applications such as controllers and planners. They acknowledge that performance may vary with different sensor configurations and noise profiles, and that the unexpected RMSE increase in one specific trajectory ("Out-of-Bounds") with larger windows may be due to sensor degradation at boundaries, suggesting that optimal window sizes are context-dependent.

The Three Ways to Guess the Shape

What the Paper Actually Did

The Trade-Off

Summary

More like this