⚡ electrical engineering

Model-Less Feedback Control of Space-based Continuum Manipulators using Backbone Tension Optimization

This paper presents a model-less feedback control framework for space-based continuum manipulators that utilizes online differential convex updates and backbone tension optimization to achieve sub-millimeter accuracy and stable actuation without relying on kinematic modeling or parameter identification.

Original authors: Shrreya Rajneesh, Rakesh Kumar Sahoo, Manoranjan Sinha

Published 2026-08-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shrreya Rajneesh, Rakesh Kumar Sahoo, Manoranjan Sinha

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

In the cramped, cluttered spaces of a spacecraft or the narrow corridors of a collapsed building, traditional robotic arms often fail. These machines, built from rigid metal links and sharp joints, struggle to navigate around obstacles without bumping into them or getting stuck. To solve this, engineers have turned to a different kind of robot: the continuum manipulator. Imagine a long, flexible spine, much like an elephant's trunk or a human finger, that can bend and twist smoothly in any direction. Because it has no hard edges, it can squeeze through tight gaps and wrap around delicate objects without causing damage. This makes them ideal for tasks like inspecting the interior of a space station or performing surgery inside the human body. However, controlling such a flexible machine is notoriously difficult. Unlike a rigid arm where the relationship between a motor's movement and the arm's position is predictable, a flexible spine changes its shape based on friction, how tightly it is pulled, and what it is touching. This unpredictability means that standard computer models, which rely on precise mathematical formulas to guess where the robot will go, often fail, leading to jerky movements or total loss of control.

Researchers at the Indian Institute of Technology Kharagpur have developed a new way to control these flexible robots that does not rely on any pre-existing mathematical model of how the robot moves. Instead of trying to calculate the complex physics of the bending spine in advance, their system learns the robot's behavior in real-time as it moves. The core of their approach is a method that constantly adjusts its understanding of the robot's movement based on what the sensors actually see. When the robot is asked to move its tip to a new spot, the computer makes a small guess, watches the result, and then immediately corrects its internal map of the robot's capabilities. This process happens so quickly that the robot can follow a path smoothly, even if the environment changes or if the robot's internal friction behaves differently than expected. The researchers tested this system in computer simulations, guiding the robot along three distinct paths: a smooth circle, a five-sided pentagon, and a square with sharp corners. In every case, the robot successfully followed the path, reaching the target with an accuracy of less than one millimeter, all without ever needing to be calibrated or taught its own mechanical properties beforehand.

A critical innovation in this work is how the system manages the internal forces within the robot's spine. These robots are typically controlled by pulling on cables, similar to how a puppeteer pulls strings. If the cables are pulled too loosely, the robot goes limp and loses its shape; if they are pulled too hard in opposing directions at the same time, the spine gets crushed and buckles. The new control system solves this by actively monitoring the tension inside the backbone itself. It ensures that the cables are always tight enough to hold the shape but not so tight that they compress the spine. This is achieved by solving a complex optimization problem at every single moment of movement, balancing the need to reach the target with the need to keep the internal structure stable. By doing this, the robot avoids the jerky, unstable movements that often plague flexible machines when they try to turn corners or change direction quickly.

The results of these simulations show that the robot can handle a wide variety of shapes with remarkable stability. When tracing a circle, the robot moved smoothly, with its tip staying within a few millimeters of the desired line. When asked to follow a pentagon, which requires the robot to bend sharply at five distinct points, the system adjusted instantly, smoothing out the corners just enough to keep the spine from breaking while still staying close to the intended path. The most difficult test was the square path, which features four extremely sharp 90-degree turns. Even here, the robot managed to track the path with high precision, deviating by only about 0.3 to 0.4 millimeters along the straight edges. The system maintained this performance by constantly updating its understanding of how the robot's cables and spine interact, effectively learning the robot's behavior on the fly. This approach proves that it is possible to control highly flexible machines in complex, confined environments without needing to know every detail of their mechanical construction beforehand.

The significance of this work lies in its ability to make these robots more reliable in unpredictable situations. In space, where a robot might need to inspect a damaged solar panel or repair a module, the environment is full of unknown obstacles and forces. A system that relies on a perfect model of the robot might fail if a cable gets slightly stuck or if the temperature changes the stiffness of the materials. By using a model-free approach that learns from real-time feedback, the robot becomes much more robust. It can adapt to changes in friction or load without human intervention. The researchers demonstrated that by combining this real-time learning with careful management of internal tension, they could achieve smooth, accurate movement that rivals more complex, model-based systems. This suggests a future where flexible robots can be deployed in hazardous or unstructured environments, such as disaster zones or deep space, with a level of autonomy and reliability that was previously difficult to achieve. The study confirms that by letting the robot teach the computer how it moves, rather than forcing the computer to guess, we can create machines that are both incredibly dexterous and surprisingly tough.

Technical Summary: Model-less Feedback Control of Space-Based Continuum Manipulators Using Backbone Tension Optimization

Problem Statement

Continuum manipulators are increasingly recognized for their intrinsic dexterity and geometric compliance, making them suitable for constrained environments such as minimally invasive surgery, search-and-rescue, and extraterrestrial inspection (e.g., on the International Space Station). However, their control in unstructured or constrained environments is hindered by several factors:

Modeling Limitations: Infinite-dimensional backbone deformation, unmodeled internal friction, and configuration-dependent stiffness make accurate model-based kinematic formulations difficult.
Control Instability: Reliance on precise geometric and mechanical parameters leads to inaccurate Jacobian predictions, artificial singularities, and unstable actuation when external disturbances or contact forces occur.
Co-activation Issues: Traditional model-less approaches often enforce only tendon tautness to avoid slack. They fail to explicitly regulate the internal axial compression (co-activation) induced by antagonistic tendon pulling, which can compromise backbone stiffness, increase curvature hysteresis, and induce buckling.

Consequently, pure model-based controllers struggle in constrained, unstructured environments, necessitating a robust, model-agnostic alternative.

Methodology

The authors propose a complete model-less feedback control framework that bypasses the need for precise kinematic or dynamic modeling. The core of the approach involves an empirically initialized Jacobian refined online through differential convex updates, integrated with a novel backbone tension optimization term.

1. Kinematic Modeling and Initialization

Surrogate Forward Kinematics: Instead of complex nonlinear elastic models, a simple surrogate geometric model (based on constant curvature approximation and affine mapping) is used solely for initialization. This provides consistent tip-displacement predictions to compute initial finite-difference Jacobian columns.
Empirical Jacobian Estimation: The task-space Jacobian ( $J$ ) is not derived analytically but is estimated online. It is initialized by perturbing actuators independently and measuring the resulting tip displacement. To prevent numerical ill-conditioning, the Jacobian columns are scaled by their Euclidean norms to create a normalized map.

2. Backbone Tension Optimization

A key innovation of this work is the explicit regulation of internal backbone axial forces ( $T_{bb}$ ) alongside tendon tensions.

Definition: For a three-tendon system, backbone tension is defined as the average of the left and right tendon tensions ( $T_{bb} = (\tau_\ell + \tau_r)/2$ ).
Objective: The controller explicitly minimizes co-activation to suppress compressive saturation and maintain longitudinal stiffness, rather than treating tension minimization merely as an energy objective. This prevents buckling during constrained maneuvers.

3. Control Architecture: Dual Convex Optimization

The control loop operates via a real-time Quadratic Program (QP) that computes actuator increments ( $\Delta y$ ) to track a desired trajectory ( $\Delta x_d$ ).

Optimization Objective: The QP minimizes a weighted sum of:
1. Task-space tracking error ( $\|J\Delta y - \Delta x\|^2$ ).
2. Tension buildup and co-activation load ( $\|\tau + K\Delta y\|^2$ ).
3. Actuator change magnitude to suppress aggressive excursions ( $\|\Delta y - \Delta y_{k-1}\|^2$ ).
Constraints: The optimization enforces:
- Slack Avoidance: Tendon tensions must remain above a minimum threshold ( $\tau \geq \tau_{min}$ ).
- Geometric Limits: Actuator displacement and total position bounds.
- Tension Limits: Upper bounds to preserve mechanical integrity.
Online Adaptation: At each control step, the Jacobian is updated using a minimal-disturbance convex program. The update minimizes the Frobenius norm of the Jacobian increment ( $\Delta J$ ) subject to matching measured displacements. The update is smoothed via exponential correction and element-clipping to prevent noise amplification and artificial singularities.

Key Contributions

Backbone Tension Optimization: The paper introduces a novel control handle that explicitly regulates internal backbone axial loading within a model-less framework. This addresses the issue of unregulated co-activation compression, which previous model-less methods ignored.
Fully Model-Agnostic Framework: The system achieves precise trajectory tracking without requiring kinematic calibration, parameter identification, or dynamic modeling of the continuum manipulator.
Dual Convex Optimization Pipeline: The integration of a real-time QP for actuator selection with a convex program for Jacobian adaptation ensures robustness against unknown environmental conditions and contact forces.

Results

The framework was validated through simulations involving three planar trajectories of increasing geometric complexity: a circle, a regular pentagon, and a square.

Trajectory Tracking: The controller demonstrated smooth convergence and sub-millimeter steady-state accuracy across all paths.
- Circle: Monotonic convergence with steady-state deviations of 1–3 mm.
- Pentagon: The manipulator aligned closely with linear edges, with local smoothing at vertices due to backbone curvature, stabilizing within 1–3 mm of segments after transitions.
- Square: Despite sharp corners, the controller achieved high precision on straight edges with errors of approximately 0.3–0.4 mm.
Tension Evolution: The system successfully maintained non-slack conditions ( $\tau \geq 0.3$ N) while exhibiting alternating dominance (one tendon tightens while the antagonist relaxes). Crucially, the backbone tension remained stable, preventing compressive saturation.
3D Deformation: 3D reconstructions confirmed that the controller generated feasible tendon updates that respected structural bending limits, producing uniform curvature for circular paths and localized curvature peaks at corners for polygonal paths.

Significance and Claims

The paper claims to establish a scalable alternative to model-dependent continuum manipulation for constrained environments. By integrating backbone tension optimization directly into the actuation-level controller, the proposed method offers:

Enhanced Stability: By regulating axial loading, the controller avoids the geometric instability and buckling associated with unregulated co-activation.
Robustness: The system operates effectively without model calibration, making it suitable for scenarios where environmental conditions (friction, contact) are unknown or variable.
Precision: The framework achieves high-precision tracking (sub-millimeter) in both continuous and non-smooth geometries, validating its utility for tasks requiring navigation through complex, obstacle-rich domains.

The authors position this work as a step toward reliable space-based and hazardous environment operations, noting that future work will extend the framework to full 3D guidance and occluded environments.