💻 computer science

Strain-Parameterized Coupled Dynamics and Dual-Camera Visual Servoing for Aerial Continuum Manipulators

This paper presents a unified strain-parameterized dynamic model for tendon-driven aerial continuum manipulators with underactuated bases and a robust dual-camera visual servoing controller, demonstrating their effectiveness through simulations and experiments on a custom-built prototype.

Original authors: Niloufar Amiri, Farrokh Janabi-Sharifi

Published 2026-03-25

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Niloufar Amiri, Farrokh Janabi-Sharifi

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 drone, but instead of having rigid, metal arms like a standard robot, it has a soft, flexible spine made of a special metal alloy. This is a Tendon-Driven Aerial Continuum Manipulator (TD-ACM). Think of it as a flying octopus or a robotic snake that can fly and grab things.

The problem? Controlling this creature is incredibly hard.

It's wobbly: Because it's soft, it bends and twists in complex ways that are hard to predict.
It's underpowered: The drone part can't move sideways directly; it has to tilt (roll and pitch) to slide over, which makes the whole system shake.
It's blind: If the drone tilts, the camera on the tip of the "snake" sees the world spin, making it hard to know where the target is.

This paper presents a solution that acts like a super-smart pilot and a nervous system combined. Here is how they solved it, broken down into simple concepts:

1. The "Virtual Camera" Trick (Fixing the Shake)

Imagine you are holding a camera on a shaky boat. If the boat tilts, the horizon in your view tilts, even if the object you are looking at hasn't moved.

The Problem: When the drone tilts to move sideways, the camera on the snake's tip spins, confusing the computer.
The Solution: The authors created a "Virtual Camera." Imagine a magical camera that sits inside the drone's brain. It takes the shaky, tilted image from the real camera and mathematically "straightens" it out, as if the drone were perfectly level. This allows the robot to ignore the drone's wobbles and focus purely on the target.

2. The "Two-Eye" System (Never Losing Sight)

Robots often lose their target if it moves out of the camera's view (like a flashlight beam).

The Problem: If the snake bends too far, the target disappears from the tip camera.
The Solution: They used two cameras.
- Eye 1 (The Tip): A high-resolution camera on the snake's nose for precise, close-up work.
- Eye 2 (The Drone Body): A wide-angle camera on the drone's body that sees everything from far away.
- How it works: If the target leaves the Tip Camera's view, the system instantly switches to the Body Camera. It uses the wide view to guide the snake back until the target is visible to the Tip Camera again. It's like having a wide-angle security camera that helps you find your lost keys before you switch to your magnifying glass to pick them up.

3. The "Smart Brain" (The Math Model)

To control a soft robot, you need a map. Old maps were too simple (assuming the snake is a straight stick) or too slow (calculating every tiny bend in real-time took too long).

The Solution: The authors built a new, ultra-fast mathematical model. They treated the snake not as a stick, but as a flexible rod that bends based on "strain" (how much it's stretched).
The Analogy: Think of it like a tightrope walker. Instead of calculating every muscle twitch, the model predicts how the whole rope will sway based on where the wind is blowing and where the walker leans. This new model is fast enough to run in real-time, allowing the robot to react instantly to wind or unexpected bumps.

4. The "Adaptive Pilot" (Handling the Unknowns)

Real life is messy. The robot's materials might stretch differently on a hot day, or the wind might change.

The Solution: They added an adaptive controller. Think of this as a pilot who doesn't just follow a script but learns as they fly. If the robot feels a sudden push from the wind or realizes its arm is heavier than expected, the controller automatically adjusts its grip and thrust to compensate. It guarantees the robot won't crash, even if the math isn't 100% perfect.

The Result: A Robust, Flying Snake

The team built a physical prototype (a small drone with a soft, bendy arm) and tested it.

In the lab: They showed it could grab a target even if the drone was shaking or the target was hidden from the tip camera initially.
In the air: They proved it could fly, tilt, and bend its arm simultaneously to reach a spot, all while keeping the target centered in its "eye."

In summary: This paper teaches a flying, soft robot how to stay steady, never lose sight of its target, and adjust to the real world's chaos, all by using a clever mix of virtual cameras, a dual-eye system, and a brain that learns on the fly. It turns a wobbly, unpredictable machine into a precise, reliable tool for future rescue missions or construction jobs in hard-to-reach places.

1. Problem Statement

The paper addresses the challenges associated with Tendon-Driven Aerial Continuum Manipulators (TD-ACMs), which combine Uncrewed Aerial Vehicles (UAVs) with compliant, lightweight continuum robots (CRs). While TD-ACMs offer superior adaptability and safety in contact-rich environments compared to rigid-link manipulators, they face three critical unresolved issues:

Modeling Complexity: Existing dynamic models either rely on simplified constant-curvature (CC) approximations (lacking accuracy under large deflections) or use fully coupled Lagrangian formulations that incur prohibitive computational costs and fail to explicitly account for the underactuation of the UAV base.
Control Uncertainties: TD-ACMs suffer from significant uncertainties, including unmodeled dynamics, material hysteresis, tendon friction, and aerodynamic disturbances. Furthermore, the underactuated nature of quadrotors creates strong coupling between translational and rotational dynamics, making robust control difficult.
Visual Servoing Limitations: Standard Image-Based Visual Servoing (IBVS) for aerial systems struggles with Field-of-View (FoV) limitations, image motion caused by UAV attitude changes (roll/pitch), and the difficulty of reconstructing the state of a deformable continuum robot without joint encoders.

2. Methodology

The authors propose an integrated framework comprising three main components:

A. Strain-Parameterized Coupled Dynamics

Unified Formulation: The authors derive a fully coupled dynamic model on the Special Euclidean group SE(3) using D'Alembert's principle.
Hybrid Modeling: The UAV is modeled as a rigid body, while the continuum manipulator is modeled using a strain-parameterized Cosserat rod approach. This avoids the infinite degrees of freedom issue by parameterizing the rod's configuration via strain variables.
Efficiency: Unlike traditional methods requiring symbolic derivations, this approach integrates the strain-based Cosserat model with the rigid-body UAV model into a unified Lagrangian Ordinary Differential Equation (ODE) framework. This eliminates computationally intensive symbolic derivations while preserving geometric accuracy and accounting for distributed body wrenches.
Actuation Mapping: The model explicitly maps the underactuated UAV inputs (thrust and torques) and tendon tensions to the generalized forces of the coupled system.

B. Dual-Camera Visual Servoing (IBVS)

Architecture: The system employs a dual-camera setup: a local camera ( $C_L$ ) mounted at the manipulator tip (Eye-in-Hand) and a global camera ( $C_G$ ) mounted on the UAV body (Eye-to-Hand).
Attitude Compensation: To mitigate image motion caused by UAV roll and pitch (which are necessary for lateral movement), a virtual camera concept is introduced on SE(3). This virtually removes attitude-induced rotations from the image space, stabilizing the visual features.
Advanced Features: Instead of simple point features, the IBVS uses a vector of advanced image features (centroid, perimeter, and orientation angles) derived from a polygon formed by four fiducial markers. This ensures global stability and decouples translational and rotational motions.
FoV Recovery Strategy: A switching logic is implemented. If features leave the local camera's FoV, the system switches to a pre-planned trajectory tracking mode using the global camera to guide the manipulator back into view. Once features are re-acquired, it seamlessly switches back to the high-precision IBVS mode.

C. Adaptive Low-Level Control

Robust Controller: A model-based adaptive sliding mode controller is designed to handle parametric uncertainties (e.g., varying stiffness, unmodeled dynamics).
Stability: The controller incorporates an adaptation law to estimate lumped uncertainties. A Lyapunov-based stability analysis proves that the tracking errors and parameter estimation errors remain bounded and converge to zero.
Hierarchical Control: For the underactuated UAV, the controller computes desired roll and pitch angles to achieve the required translational motion, ensuring the coupled system tracks the reference trajectory generated by the high-level visual servo.

3. Key Contributions

Unified Dynamic Model: A geometrically exact, strain-parameterized coupled dynamic model on SE(3) that balances accuracy and computational efficiency, avoiding the need for symbolic derivations.
Dual-Camera IBVS Framework: A novel visual servoing scheme that mitigates FoV loss, compensates for UAV attitude-induced image motion, and leverages the complementary advantages of local and global cameras.
Adaptive Control with Stability Guarantees: A robust low-level controller that formally guarantees stability in the presence of modeling uncertainties and strong UAV-CR coupling.
Experimental Validation: Extensive validation using a custom-built TD-ACM prototype, demonstrating performance in both ground-based and fully coupled aerial scenarios.

4. Results

Computational Efficiency: In simulations, the proposed dynamic formulation achieved an average computational time of 0.0339 s per iteration, compared to 0.0460 s for a baseline coupled Lagrangian model (a 26.3% reduction).
Simulation Performance:
- The system successfully guided the manipulator tip to a target with a normalized Root Mean Square Error (RMSE) decreasing from 0.2 to 0.01.
- The dual-camera switching strategy successfully recovered from an "out-of-FoV" condition, transitioning smoothly from trajectory tracking to IBVS.
Experimental Validation:
- Test A (Ground-based): Demonstrated convergence of feature errors and stable velocity regulation.
- Test B (FoV Recovery): Successfully recovered the target from outside the local camera's view using the global camera and pre-planned trajectories.
- Test C (Aerial Coupled): The most complex test involving full UAV-CR coupling. Despite environmental disturbances (vibrations, downwash) and strong dynamic coupling causing initial oscillations, the system stabilized, and feature errors decreased monotonically. The system maintained steady flight while holding the desired configuration.

5. Significance

This work represents a significant advancement in the field of aerial robotics by providing a computationally tractable, fully coupled dynamic model for TD-ACMs, which has been a major bottleneck for real-time control. The proposed dual-camera visual servoing solves the critical issue of FoV limitations and attitude-induced instability, enabling reliable operation in unstructured environments. The experimental validation confirms that the framework is robust enough for real-world deployment, paving the way for complex aerial manipulation tasks such as inspection, maintenance, and interaction in GPS-denied or cluttered environments.