A Stochastic Framework for Continuous-Time State Estimation of Continuum Robots

arXiv:2510.01381 · cs.RO · Submitted 2025-10-01 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "A Stochastic Framework for Continuous-Time State Estimation of Continuum Robots".

Dev: Here's the gist of the paper. State estimation techniques for continuum robots (CRs) typically involve using computationally complex dynamic models, simplistic shape approximations, or are limited to quasistatic methods <ref:2510.01381#pg2>.

Rosa: Inspired by a factor-graph optimization paradigm, this work introduces a continuous-time stochastic state estimation framework for continuum robots <ref:2510.01381#pg3>.

Dev: Pruning redundant reasoning tokens This work addresses the gap in probabilistic state estimation methods by proposing a physically motivated Gaussian-process prior defined over both space and time, unifying continuous-time

37: with continuous arclength

36: estimation into a single framework <ref:2510.01381#pg4>.

Rosa: The method is the first stochastic state estimator for continuum robots that is continuous in time <ref:2510.01381#pg5>.

Paper summary: Rosa: So, we're looking at this paper today called "A Stochastic Framework for Continuous-Time State Estimation of Continuum Robots".

Dev: It looks like they're tackling the problem of figuring out where a continuum robot is in real time.

Rosa: Exactly. Usually, these estimation techniques are either using really complex models, or they use very simple shapes, or they only work well when things are mostly steady and slow.

Taro: That sounds like it can be sensitive to unexpected things happening with the robot while you're driving or whatever you're doing.

Dev: Right. This paper introduces a continuous-time stochastic state estimation framework, inspired by a factor-graph optimization approach, to fix those limitations.

Rosa: So what’s the core claim here? What are they trying to achieve that the old methods couldn't?

Dev: They propose a physically motivated Gaussian process prior that covers both space and time. This unifies continuous-time estimation with continuous arclength estimation into one single framework, which is pretty neat Rosa

Rosa: That sounds like it solves a major gap in how we track these robots. What does this mean practically for someone listening who just wants to know if the system is working?

Dev: The main thing they claim is that this is the first stochastic state estimator for continuum robots that's continuous in time, which is pretty significant Taro It also supports handling high-rate and asynchronous measurements with minimal computational overhead, so it should be efficient Rosa

Taro: I'm interested in how they handle the noise. They simplify a dynamic Cosserat rod model by injecting white noise onto the second order derivative terms, specifically using 'white-noise-on-acceleration' assumptions Dev That’s a specific way they inject that uncertainty into the system.

Dev: Yeah, and they construct their prior factors as a sum of binary factors in both time and space. This creates this full factor graph where nodes represent states at specific arclengths and times Rosa The solver uses a Gauss-Newton solver to minimize the total cost function, which means it's an iterative process that converges when the perturbation vector magnitude is sufficiently small Taro

Rosa: Okay, so they are using this structure—pose, velocity, and strain—to get estimates for all three. Can we get some concrete results on how good this is? What did they test it with?

Dev: They validated their approach through both simulation and experimental evaluation using pose, strain, and velocity measurements on a tendon-driven continuum robot Rosa In simulation results, they show an average tip error of zero point one nine five percent for the robot's trajectory Taro

Rosa: That error rate sounds pretty good for something dynamic. But what about when things get messy in the real world? How does this framework handle data dropout or unmodelled external forces?

Dev: The results highlight the method’s ability to capture dynamic effects in the robot's motion, handle data dropout, and benefit from sensor fusion Taro Experimentally, they found that when using pose measurements, the average tip error stays below two percent for all trajectories Rosa <ref:2510.01381#pg2>

Paper summary: Taro: It sounds like it's robust enough for real-world testing on a tendon-driven robot. But what about the computational side? How does this framework scale when you add more complexity?

Dev: The runtime scales linearly with respect to the number of time nodes K, and cubically with respect to the number of spatial nodes N Taro That cubic scaling with spatial nodes is something we'll need to watch out for in real-time applications.

Rosa: So, if we put this into practice, what’s the big picture here? What does this paper actually mean for how people use estimation techniques in robotics?

Dev: It opens the door to further cross-domain adaptation from works established in mobile robotics Taro This framework provides a flexible foundation for adding other sensors like cameras, IMUs, or strain gauges later on Rosa But they do flag that future work should explore computationally efficient methods for computing the covariance in real time Dev

Rosa: And they mention that the solver remains linear in time with respect to K, which is important if you need it to run fast enough for real-time applications Dev It's a solid foundation for building more complex systems.

Taro: I think the main thing here is moving from just fitting a shape to actually estimating the true state of motion over space and time simultaneously. That's what this paper delivers Rosa

Dev: Yeah, it’s about using that factor-graph optimization idea to get a continuous picture of the robot's pose, velocity, and strain all at once. It’s a lot to handle in real time Taro

Rosa: So, that's where we are with the paper "A Stochastic Framework for Continuous-Time State Estimation of Continuum Robots". It’s a framework that uses continuous-time kinematics corrupted by a white noise Gaussian process to estimate the mean and covariance for pose, velocity, and strain continuously in time or space Dev

Taro: It’s an interesting piece because it addresses the gap in existing continuum robot estimation methods by unifying continuous-time and continuous arclength estimation into one framework Rosa

Dev: The authors show that with their specific assumptions about white noise on acceleration, they can achieve a result where the error falls with the square of the sampling interval when we look at simulation Taro It’s a concrete detail showing how their assumptions translate to performance.

Rosa: So, this paper suggests that for future work, we should focus on making that covariance computation faster for real-time use Dev It gives us a solid way to start incorporating more sensors onto these continuum robots later Taro

Dev: That's the direction they point toward. It’s a framework built on factor graphs that lets us handle the continuous nature of these robots better than we could before Rosa

Conclusion: Rosa: It’s just a framework for getting a better picture of where these long robots are located over time and space. The authors built it using some ideas from factor graphs to make sure the estimation is continuous.

Dev: Yeah, it tackles the problem that old methods had, which were often too slow or only worked well when things were very steady. It’s about creating a system that can track movement even when things are changing rapidly.

Taro: I see why it's important because it allows us to handle the messy reality of robot motion, not just the simple math models we used before.

Rosa: It seems like this is really about giving these robots a way to know their own state—their position, their speed, and even how they are stretching—all at the same time.

Dev: Exactly. And what they do with the training signal is clever; they build factors both in time and in space so it can handle those high-speed measurements we get from sensors.

Taro: That continuous nature of it is what makes me interested, because it means if we have a sensor reading at one spot, we can smoothly connect that data to another spot later on.

Rosa: It’s not just about getting a number; it’s about building a model that evolves as the robot moves. It's giving these robots a kind of memory for their path.

Dev: And the performance numbers they showed, like that small tip error in simulation, show that this approach is actually doing something useful outside of just theory.

Taro: Yeah, and it shows it can handle those real-world bumps and sensor dropouts better than what we've seen with simpler methods.

Rosa: So, the big implication here is that we can start building more complex systems because we have a solid way to track these flexible robots in continuous motion.

Dev: It gives us a flexible foundation for adding other things later on, like cameras or strain gauges, which is huge for real-world applications.

Taro: The next thing we need to figure out is how fast this entire estimation process can actually run when you put it into a system that needs to react instantly.

Rosa: That’s what we'll be talking about next—how this framework handles the computational demands of keeping up with real-time movement.

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

University of Toronto Robotics Institute

cs.RO

Submitted: 2025-10-01

Updated: 2026-11-13

Comments: 20 pages, 15 figures. Accepted to IEEE Transactions on Robotics. Notable changes from v1 include switching to a GTSAM back-end, adjusting the spatial factor definition, and the inclusion of a sensitivity study

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 75/100

The gist: State estimation techniques for continuum robots (CRs) typically involve using computationally complex dynamic models, simplistic shape approximations…

Key concepts

Factor-Graph Optimization
This approach models the estimation problem as a graph where nodes represent states at specific points (in space and time) and edges represent constraints or factors derived from measurements. The goal is to find the most probable state configuration by minimizing a total cost function across all these connected nodes.
Gaussian-Process Prior
This is a physically motivated prior defined over both the robot's spatial configuration and its evolution over time. It helps constrain the possible states, making the estimation more robust by assuming certain smoothness and continuity in how the robot moves through space and time.
White-Noise Assumptions
The method simplifies complex dynamic models by injecting white noise into second-order derivative terms (like acceleration). This allows for a computationally tractable prior construction, enabling efficient handling of high-rate measurements without needing overly complex physical simulations.

Terminology

Summary

State estimation techniques for continuum robots (CRs) typically involve using computationally complex dynamic models, simplistic shape approximations, or are limited to quasistatic methods <ref:2510. p. 0

Where the paper says it

1 A Stochastic Framework for Continuous-Time State Estimation of Continuum Robots Spencer Teetaert, Graduate Student Member, IEEE, Sven Lilge, Member, IEEE, Jessica Burgner-Kahrs, Senior Member, IEEE, Timothy D. Barfoot, Fellow, IEEE Abstract—State estimation techniques for continuum robots (CRs) typically involve using computationally complex dynamic models, simplistic shape approximations, or are limited to quasistatic methods.

01381#pg2>. Inspired by a factor-graph optimization paradigm, this work introduces a continuous-time stochastic state estimation framework for continuum robots <ref:2510. p. 0
Where the paper says it

These limitations can be sensitive to unmodelled disturbances acting on the robot. Inspired by a factor-graph optimization paradigm, this work introduces a continuous-time stochastic state estimation framework for continuum robots.

01381#pg3>.

Pruning redundant reasoning tokens

This work addresses the gap in probabilistic state estimation methods by proposing a physically motivated Gaussian-process prior defined over both space and time, unifying continuous-time [37] with continuous arclength [36] estimation into a single framework <ref:2510.01381#pg4>. The method is the first stochastic state estimator for continuum robots that is continuous in time <ref:2510. p. 2

Where the paper says it

To the best of our knowledge, this work represents the first stochastic state estimator for continuum robots that is continuous in time. The proposed method supports efficient handling of high-rate and asynchronous measurements with minimal computational overhead.

01381#pg5>. It supports efficient handling of high-rate and asynchronous measurements with minimal computational overhead <ref:2510. p. 2
Where the paper says it

To the best of our knowledge, this work represents the first stochastic state estimator for continuum robots that is continuous in time. The proposed method supports efficient handling of high-rate and asynchronous measurements with minimal computational overhead.

01381#pg5>. The result contains an estimate of the mean and covariance for the robot’s pose, velocity, and strain, each of which can be interpolated continuously in time or space <ref:2510. p. 0
Where the paper says it

By using a simple robot model paired with high-rate sensing, we show adaptability to unmodelled external forces and data dropout. The result contains an estimate of the mean and covariance for the robot’s pose, velocity, and strain, each of which can be interpolated continuously in time or space.

01381#pg5>.

How the training signal is built

The estimation framework is constructed by defining a state representation that includes a pose T (s, t) ∈ SE(3), a bodycentric strain ϵ(s, t) ∈ R6, and body-centric velocity of the robot ϖ(s, t) ∈ R6 <ref:2510.01381#pg4>. The prior is approximated through the construction and solution of a factor-graph optimization problem <ref:2510. p. 2

Where the paper says it

(3) This work does not propose such a closed-form representation. Rather, we will approximate the prior through the construction and solution of a factor-graph optimization problem, relying heavily on analogy to the time and space systems presented in [36], [37].

01381#pg4>. This involves constructing time factors relating the state at a specific point in time, tk, to the state at some later point in time, tk+1 <ref:2510. p. 3
Where the paper says it

Afterwards, we will show that we can interpolate between these states in time or space, resulting in a continuous state representation. 1) Time Factors: We begin by constructing a series of factors that relate the state at a specific point in time, tk, to the state at some later point in time, tk+1.

01381#pg5>. Similarly, arclength factors relate the state between states at two points in arclength, sn and sn+1 <ref:2510. p. 3
Where the paper says it

(12) 2) Arclength Factors: We continue by constructing factors for the arclength parameter between states at two points in arclength, sn and sn+1, along the backbone of the robot. Beginning with the arclength dynamics of our state variables, ∂T ∂s = ϵ ∧T , ∂ϵ ∂s ≈ w2, ∂ϖ ∂s ≈ ϵ ⋏ϖ + w3.

01381#pg5>.

Where the gains come from

The core of the method involves a significant simplification of a dynamic Cosserat-rod model by injecting white noise onto the second-order derivative terms, specifically using ‘white-noise-on-acceleration’ (and the analogous ‘white-noise-on-strain-rate’) assumptions <ref:2510.01381#pg5>. The prior factors are constructed as a sum of binary factors in both time and space, leading to a full factor graph where nodes represent states at specific arclengths and times <ref:2510. p. 5

Where the paper says it

A grid is assembled using N points in space and K points in time. The nodes of this grid represent states at specific arclengths and times that will be explicitly estimated by the estimator.

01381#pg5>. The solver uses a Gauss-Newton solver to minimize the total cost function, resulting in an iterative process that converges when the perturbation vector magnitude is sufficiently small <ref:2510. p. 7
Where the paper says it

The state is updated with the perturbation, T op(s, t) ← exp(δt ∗ (s, t) ∧ )T op(s, t), (43a) ϵop(s, t) ← ϵop(s, t) + δϵ ∗ (s, t), (43b) ϖop(s, t) ← ϖop(s, t) + δϖ∗ (s, t) (43c) and the process is repeated until convergence. The optimization process is terminated when the magnitude of the perturbation vector is sufficiently small, ||δx ∗ || < ϵtol, (44) or a maximum number of iterations is reached.

01381#pg9>.

Validation and performance

The method was validated through both simulation and experimental evaluation using pose, strain, and velocity measurements on a tendon-driven continuum robot <ref:2510.01381#pg5>. Simulation results show an average tip error of 0. p. 9

Where the paper says it

We also present the root mean square error (RMSE) of the robot tip (and body) pose, which we compute as dRMSE,pos = 1 L vuut 1 N X N n=1 ||p˜n − pn||2 2 , (50a) dRMSE,rot = vuut 1 N X N n=1 ||ln(R˜ nR −1 n ) ∨ ||2 2 . (Simulation) Robot Extension The estimator is evaluated on the extensible robot simulation trajectory ‘Extensible’.

195% for the extensible robot trajectory <ref:2510. p. 8
Where the paper says it

We simulate a three-segment TDCR that is fully extensible, which is to say, each of the segments can span from 0 to a full length of 1 3 and back to 0. The robot has a length of 1 and has a trajectory of duration 1, both unitless values.

01381#pg10>. The estimator demonstrates its ability to capture dynamic effects in the robot’s motion, handle data dropout, and benefit from sensor fusion <ref:2510. p. 2
Where the paper says it

We validate our approach through both simulation and experimental evaluation, using pose, strain, and velocity measurements on a tendondriven continuum robot. Our experimental results highlight the method’s ability to capture dynamic effects in the robot’s motion, handle data dropout, and benefit from sensor fusion.

01381#pg10>. Experimental results show that the average tip error remains below 2% for all trajectories when pose measurements are used <ref:2510. p. 2
Where the paper says it

We validate our approach through both simulation and experimental evaluation, using pose, strain, and velocity measurements on a tendondriven continuum robot. Our experimental results highlight the method’s ability to capture dynamic effects in the robot’s motion, handle data dropout, and benefit from sensor fusion.

01381#pg10>. The runtime scales linearly with respect to the number of time nodes K and cubically with respect to the number of spatial nodes N <ref:2510. p. 7
Where the paper says it

The Cholesky factor is lower triangular, with non-zero entries bounded about the diagonal. The band scales with N, the number of spatial nodes in our state, leading to a solve complexity of O(N3K), where K is the number of time steps in our estimation window.

01381#pg10>.

Conclusion

In this work, we presented a fully stochastic, continuous-time estimation framework for continuum robots by extending the works of [36] and [37] to the domain of continuum robot trajectory estimation (i.e., space and time) <ref:2510. p. 0

Where the paper says it

1) Non-probabilistic Methods: Non-probabilistic estimation methods usually employ optimization approaches to fit kinematic, static, or dynamic models to discretely sensed quantities, such as pose or strain. arXiv:2510.

01381#pg10>. By formulating the continuous-time trajectory estimation of a continuum robot as a factor-graph optimization problem, we hope to open the door to further cross-domain adaptation from works established in mobile robotics <ref:2510. p. 0
Where the paper says it

These limitations can be sensitive to unmodelled disturbances acting on the robot. Inspired by a factor-graph optimization paradigm, this work introduces a continuous-time stochastic state estimation framework for continuum robots.

01381#pg10>. This framework provides a flexible foundation for the addition of other sensors such as cameras, IMUs, or strain gauges <ref:2510. p. 5
Where the paper says it

(25) While we have presented measurement models for pose and gyroscope sensors, the framework is general enough to include other sensor modalities provided they have a sensor model that relates the measurement to the state variables in x(s, t) and a known Jacobian. Fiber Bragg Grating Sensor [36], strain gauges [40], and stereo cameras [37] have been included in similar frameworks in past works.

01381#pg10>. It is left to future work to explore computationally efficient methods for computing the covariance in real-time <ref:2510.01381#pg10>. The proposed solver remains linear in time (K), a fact that is important for real-time applications <ref:2510. p. 7
Where the paper says it

Including the cost of matrix construction, the total time complexity of computing the posterior is O(N3K + M) where M is the number of measurements included. The proposed solver remains linear in time (K), a fact that is important for real-time applications.

01381#pg10>. The authors provide an open-source implementation of this project to the research community at https://github. p. 2
Where the paper says it

Our experimental results highlight the method’s ability to capture dynamic effects in the robot’s motion, handle data dropout, and benefit from sensor fusion. An open-source implementation of this project is made publicly available to the research community at https://github.

com <ref:2510. p. 0
Where the paper says it

1) Non-probabilistic Methods: Non-probabilistic estimation methods usually employ optimization approaches to fit kinematic, static, or dynamic models to discretely sensed quantities, such as pose or strain. arXiv:2510.

01381#pg10>.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Continuous-time state estimation for continuum robots allows AI systems to operate in real-time by fusing asynchronous, high-rate sensor data with a physically motivated Gaussian process prior defined over both space and time. This enables the system to produce an estimate of the mean and covariance for pose, velocity, and strain that can be interpolated continuously in time or space during estimation.

  2. The framework supports efficient handling of high-rate and asynchronous measurements with minimal computational overhead, achieving solve runtimes that are approximately real-time when only the mean is computed. This speed is critical for applications requiring rapid decision-making in dynamic environments.

  3. The system can maintain a reasonable estimate of the robot's state despite unmodelled external forces by incorporating noise terms defined as GP white noise, such as 'white-noise-onacceleration' or 'white-noise-onstrain-rate'. This provides adaptability to unmodelled disturbances and data dropout without requiring explicit, complex models for every external interaction.

  4. The AI system can perform robust state estimation even when measurements are lost in certain regions of the robot's workspace, such as when the robot tip leaves the sensor workspace. The proposed method is able to generate a reasonable estimate while accounting for the increased uncertainty in its confidence bounds during these dropout scenarios.

  5. The framework enables tuning of process noise matrices Qi and measurement noise terms Ri based on sensor characteristics and expected uncertainty along different state dimensions. This allows users to tailor the estimator's behavior, such as setting low values for shearing strain uncertainties to ensure high certainty in initial shear strain estimates.

Abstract

State estimation techniques for continuum robots (CRs) typically involve using computationally complex dynamic models, simplistic shape approximations, or are limited to quasi-static methods. These approaches can be sensitive to unmodeled disturbances acting on the robot. Inspired by a factor-graph optimization paradigm, this work introduces a continuous-time stochastic state estimation framework for continuum robots. We introduce factors based on continuous-time kinematics that are corrupted by a white noise Gaussian process (GP). By using a simple robot model paired with high-rate sensing, we show adaptability to unmodeled external forces and data dropout. The result contains an estimate of the mean and covariance for the robot's pose, velocity, and strain, each of which can be interpolated continuously in both space and time. This same interpolation scheme can be used during estimation, allowing for inclusion of measurements that are asynchronous to the estimated states. Our method's inherent sparsity leads to a linear solve complexity with respect to time and interpolation queries in constant time. We demonstrate our method on a CR with gyroscope and pose sensors, highlighting its versatility in real-world systems.

Sources

Related papers