💻 computer science

Bayesian Continuum Robot Dynamics and State Estimation

This paper presents a Bayesian framework that approximates Cosserat rod dynamics by reformulating inertial and damping effects as equivalent loads, enabling accurate joint state estimation and external load inference for continuum robots during dynamic motions where traditional quasi-static methods fail.

Original authors: James M. Ferguson, Tucker Hermans, Alan Kuntz

Published 2026-09-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: James M. Ferguson, Tucker Hermans, Alan Kuntz

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

Soft, flexible robots that bend and twist like living tissue are opening new doors in medicine, capable of slipping through narrow passages where rigid machines cannot go. These "continuum" robots, often shaped like long, slender tubes, are ideal for delicate tasks inside the human body, such as navigating the complex curves of the heart or the airways. For years, engineers have relied on simplified models to predict how these robots move. These models work well when the robot moves slowly and steadily, treating the machine as if it were frozen in time, balancing forces at a single moment. However, the real world is rarely static. When these robots move quickly, or when they are large and light enough to be easily tossed by their own momentum, the old models break down. They fail to account for the physics of motion itself—the way inertia carries a moving object forward and how damping slows it down. Without understanding these dynamic forces, a robot might overshoot its target, vibrate uncontrollably, or misjudge the pressure it is applying to a patient.

A team of researchers has developed a new way to track and predict the movement of these flexible machines, one that embraces the chaos of motion rather than ignoring it. Instead of pretending the robot is always still, their method treats the robot's shape, speed, and the forces acting upon it as a continuous story unfolding over time and space. They built a mathematical framework that acts like a high-fidelity simulator when no sensors are watching, and a precise detective when data is available. By combining the laws of physics that govern how rods bend and twist with modern probability theory, they created a system that can estimate exactly where the robot is, how fast it is moving, and what forces are pushing or pulling on it, even when the robot is swinging wildly.

The core of this new approach is a shift in perspective. Previous methods often tried to solve the robot's movement by looking at each moment in isolation, assuming that if the robot is moving fast, the forces are just a minor correction to a slow, steady state. The researchers found that this assumption is dangerous. When a robot accelerates, its own mass creates a force that resists that change, much like a heavy suitcase feels heavier when you try to lift it quickly. In their new model, they treat this resistance not as a complication, but as a load, just like gravity or a hand pushing on the robot. By rewriting the equations of motion to include these inertial effects as standard forces, they could use a powerful statistical tool called a factor graph. This tool connects the robot's past, present, and future states into a single, cohesive picture, allowing the system to learn from every piece of data it receives.

To test their idea, the team first ran the model in a computer simulation without any real-world sensor data, essentially asking the math to predict how a flexible rod would behave if it were struck and then left to vibrate on its own. The results matched a well-known, highly detailed physics simulation almost perfectly, but the new method used far fewer calculation steps to get there. More importantly, it provided a measure of confidence for every prediction, telling the user not just where the robot was, but how sure the system was about that location. This is crucial for safety; a robot that knows it is uncertain can slow down or ask for more information, whereas a robot that is confidently wrong can cause damage.

The researchers then put the system to the test against a robot that was actually moving. They used data from a tendon-driven robot, a type of soft machine controlled by cables that pull on its body, similar to how muscles pull on bones. In one experiment, they tried to guess the force the robot was exerting on its environment just by watching how its cables were pulled and where its tip was located. The old, slow-motion models failed spectacularly here. Because they could not account for the robot's own momentum, they mistook the vibrations caused by the robot's speed for external forces, leading to errors that were five times larger than the new method. The new model, however, correctly separated the robot's own motion from the forces it was applying, keeping its estimates accurate even as the robot swung and bounced.

In a second test, the team asked the system to figure out the robot's shape using only the tension in its control cables, without any cameras watching its body. The old model, again ignoring the dynamics, predicted a shape that drifted far from reality, unable to explain why the robot was vibrating. The new model, which understood the physics of the swing, tracked the robot's true shape with much higher precision. When they added a special fiber-optic sensor that could feel the bending of the robot's spine, the accuracy improved even further, proving that the system can seamlessly blend different types of information. The entire process ran in real time, taking roughly 40 to 55 milliseconds to update its guess for every fraction of a second of movement.

The work demonstrates that for flexible robots, especially those that are large or move quickly, ignoring the physics of motion is no longer an option. The new method provides a single, unified way to simulate these machines and track them in the real world, offering a level of accuracy and self-awareness that was previously out of reach. While the system is currently slightly slower to compute than the old, simplified models, the trade-off is necessary for safety and precision in dynamic environments. As these robots move from controlled labs to complex, unpredictable tasks, the ability to understand and predict their motion in real time will be the difference between a successful procedure and a failure.

Problem Statement
Continuum robots, valued for their compliance and ability to operate in confined environments, are traditionally modeled using quasi-static assumptions. These models assume that inertial effects are negligible compared to actuation and external loads, allowing the robot's shape to be estimated as an equilibrium state at a single moment in time. While effective for slow, damped, or low-amplitude motions, these approximations fail when significant inertial effects are present, such as in larger, more compliant manipulators or during dynamic tasks. Existing dynamic modeling approaches often rely on Partial Differential Equation (PDE) solvers or constant curvature models that do not inherently account for uncertainty or allow for the fusion of online backbone observations. Furthermore, recent Bayesian state estimation methods for continuum robots have utilized kinematic motion priors (e.g., white-noise-on-acceleration) which, while efficient, do not capture the physics of actuation, inertia, or damping, leading to inaccuracies when measurements are sparse or motion is dynamic.

Methodology
The authors propose a Bayesian factor graph framework that approximates Cosserat rod dynamics to perform spatiotemporal state estimation. The core innovation is reformulating the dynamic balance equations to retain the algebraic structure of the static balance used in prior quasi-static work. By treating inertia and damping as equivalent applied loads, the dynamic equation is rewritten as:
$K\frac{\partial \epsilon}{\partial s} = \epsilon^\wedge T K(\epsilon - \epsilon_0) - f_{tot}$
where the total distributed wrench $f_{tot}$ includes external loads, tendon actuation, gravity, and an inertial wrench term derived from the D'Alembert inertial load and linear viscous damping.

The system is discretized into a factor graph spanning both space (along the robot's backbone) and time. Key components include:

Spatial and Temporal Kinematics: Pose evolution is modeled using Piecewise-Linear Magnus integration, which provides higher numerical accuracy than piecewise-constant integration by accounting for the non-commutativity of rates at interval endpoints.
Compatibility Factor: Enforces the integrability condition linking spatial strain and temporal velocity, ensuring the estimated states correspond to a realizable rod motion.
Momentum Balance Factor: Adapts the static wrench balance to include the full inertial and damping loads.
Actuation and Inertial Factors: Tendon actuation is modeled via tension variables and routing geometry, while the inertial wrench is computed using a second-order backward differentiation formula (BDF2) for stability and accuracy in stiff systems.
Optimization: The graph is solved as a fixed-lag smoother over a sliding time window. Older states are marginalized into a history prior, allowing the system to maintain a joint distribution over shape, velocity, strain, actuation, and external loads while bounding computational cost.

The framework supports heterogeneous sensing, fusing data from tip trackers, tendon tension sensors, and Fiber Bragg Grating (FBG) shape sensors without altering the underlying dynamics model.

Key Contributions

Dynamic Factor Graph Formulation: The paper presents the first factor graph formulation of Cosserat rod dynamics that explicitly models actuation, inertia, and damping while retaining the algebraic form of static balance.
Unified Framework: The approach serves a dual purpose: as a stochastic forward simulator when no measurements are available, and as a principled Bayesian state estimator when measurements are present.
Uncertainty Quantification: Unlike PDE-based solvers, this method reports uncertainty (covariance) at every timestep, indicating the reliability of the estimated shape and states.
Modular Sensing: The factor graph structure allows for the seamless integration of diverse sensor modalities (e.g., tip position and internal strain) as additional factors.

Results
The approach was validated through simulation and real-world experiments:

Forward Simulation: In a simulation without measurements, the method reproduced the results of a benchmark Cosserat rod dynamic solver [11] with an RMS tip position error of 2.80 mm, using fifteen times fewer spatial nodes. It successfully captured the stochastic nature of the motion.
Tip Force Estimation: In a simulation with noisy tip position and tension measurements, the dynamic estimator accurately inferred tip forces where a quasi-static estimator failed. The quasi-static model, lacking an inertia term, misinterpreted inertial oscillations as external loads, resulting in a force error roughly five times higher than the proposed method (733 mN vs. 152 mN). Crucially, the quasi-static model remained "confident" (tight uncertainty bounds) despite being incorrect, whereas the dynamic model provided calibrated uncertainty.
Real-World Shape Estimation: Using an open-source dataset of a tendon-driven continuum robot undergoing dynamic oscillatory motion, the dynamic model reduced the RMS tip position error from 83.7 mm (quasi-static) to 48.4 mm. Incorporating FBG shape sensor data further reduced the error to 25.6 mm. In all cases, the true tip position fell within the estimated $2\sigma$ uncertainty bounds.
Computational Cost: The dynamic solver required approximately 40–55 ms per step, compared to 2.1–7.5 ms for the quasi-static solver, reflecting the trade-off between modeling complexity and accuracy.

Significance and Claims
The paper claims that for continuum robots where inertial effects are significant, dynamic modeling is necessary for accurate state estimation and load inference. The proposed method addresses the limitations of quasi-static models by correctly attributing tension changes to acceleration rather than elastic deformation. By posing dynamics as an inference problem on a factor graph, the approach provides a single probabilistic framework that offers calibrated uncertainty, which is critical for downstream tasks like control and safety monitoring. The authors note that while the method is computationally more expensive than quasi-static approaches, it is essential for larger or faster manipulators where inertial forces are comparable to external loads. The work establishes a foundation for future applications in model-predictive control and handling distributed or unknown loads.

More like this