⚡ electrical engineering

Contact-Aware Planning and Control of Continuum Robots in Highly Constrained Environments

This paper presents a contact-aware planning and control framework for continuum robots that evaluates and penalizes hazardous interactions while permitting benign contact, enabling reliable navigation with 100% success in reaching targets within highly constrained anatomical environments.

Original authors: Aedan Mangan, Kehan Long, Ki Myung Brian Lee, Miheer Potdar, Nikolay Atanasov, Tania K. Morimoto

Published 2026-04-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Aedan Mangan, Kehan Long, Ki Myung Brian Lee, Miheer Potdar, Nikolay Atanasov, Tania K. Morimoto

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 you are trying to guide a very long, flexible, and delicate worm through a maze made of fragile, squishy glass tubes. This is essentially what doctors face when they try to navigate a catheter (a thin medical tube) through a patient's blood vessels to treat a stroke. The goal is to get the tip of the tube to a specific spot deep inside the brain without poking a hole in the glass walls.

This paper presents a new "brain" for a robotic worm (called a continuum robot) that helps it solve this maze safely and efficiently. Here is how it works, broken down into simple concepts:

1. The Problem: The "Too Tight" Maze

Traditional robots are like rigid sticks; they break if they hit a wall. Soft robots are like worms; they can bend and squeeze. But there's a catch:

Too much contact is bad: If the tip of the robot bumps into the vessel wall, it could cause a tear (like poking a hole in a water balloon).
Too little contact is hard: Sometimes, the robot needs to lean against the wall to push itself forward, just like a climber uses a rock to push up.
The Trap: If the robot leans the wrong way, it can get "stuck" in a position where it can't move anymore (a singularity), or it might lose control.

2. The Solution: A "Contact-Aware" GPS

The researchers created a planning system that acts like a super-smart GPS. Unlike a normal GPS that just says "avoid the wall," this one understands the quality of the contact.

The "Good" vs. "Bad" Contact: Imagine the robot has a "safety zone."
- Bad Contact: The very tip of the robot touching the wall. The GPS says, "NO! That's dangerous!" and reroutes.
- Benign Contact: The middle of the robot's body leaning against the wall to help it turn. The GPS says, "Okay, that's fine, it helps us move."
The "Traffic Light" System: The planner divides the maze into different zones. In some areas, it tells the robot to just go straight. In tight corners, it tells the robot, "You need to wiggle your tail and push off the wall carefully."

3. The "Magic Map" (Task-Space Partitions)

To make the planning fast, the robot doesn't look at the whole maze at once. It breaks the map into "neighborhoods" based on how hard it is to get to the goal.

Simple Neighborhoods: If you are in a straight hallway, the robot just moves forward.
Complex Neighborhoods: If you are near a tricky fork in the road, the robot knows it needs to use more complex moves (like bending its body in specific ways) to get through.
The Analogy: Think of it like a hiker. If they are on a flat trail, they just walk. If they are near a cliff edge, they switch to "climbing mode" and move very carefully. The robot does the same thing automatically.

4. The "Self-Correcting" Driver (Control)

Even with a perfect map, real life is messy. The robot's body might stretch slightly differently than expected, or the blood vessel might be in a slightly different spot than the scan showed.

The Feedback Loop: The robot has a tiny sensor at its tip (like a blind person's cane). As it moves, it constantly checks: "Am I where I thought I would be?"
The Correction: If it drifts even a millimeter, the robot's "brain" instantly adjusts the motors to steer it back on track. It's like a self-driving car that gently corrects its steering wheel every second to stay in the lane, even if the road is bumpy.

5. The Results: A Perfect Score

The team tested this system on 3D-printed models of human aortic arches (the big curve at the top of the heart) taken from real patient scans.

Success Rate: In every single test (100%), the robot reached the target.
Safety: It never let the dangerous tip touch the wall.
Precision: It stayed incredibly close to the planned path (within about 1-2 millimeters, which is the width of a pencil lead).
The "What If" Test: When they turned off the "safety rules" (letting the tip touch the wall), the robot got stuck or failed. This proved that the "contact-aware" rules were the secret sauce.

The Big Picture

This paper is about teaching a soft robot to be tactful. It knows when to be gentle, when to push, and when to avoid contact entirely. By combining a smart map that understands the dangers of the maze with a driver that constantly corrects its path, they have created a system that could one day help doctors perform life-saving stroke treatments with less risk and more precision.

In short: They built a robot that knows how to dance through a minefield without stepping on the mines, using its body to lean on the safe spots while keeping its head (the tip) perfectly clear.

1. Problem Statement

Continuum (soft) robots are ideal for navigating confined, deformable, and fragile environments like human vasculature (e.g., the aortic arch). However, these environments present a unique challenge: contact is unavoidable.

The Dilemma: While controlled contact can aid navigation, unfavorable contact (specifically at the robot's distal tip or end-of-continuum-segment) can cause vessel perforation, induce kinematic singularities, degrade manipulability, or cause hardware failure.
Limitations of Prior Work: Existing planners typically either avoid all contact (ignoring the necessity of contact for navigation) or exploit contact indiscriminately. Furthermore, they often lack integration with closed-loop control, meaning planned trajectories may be kinematically infeasible or unsafe when executed on real hardware due to model inaccuracies and environmental uncertainty.
Goal: The authors aim to develop a unified framework that distinguishes between "benign" (safe) and "hazardous" (dangerous) contact, generates kinematically feasible trajectories, and executes them via robust closed-loop control in real anatomical models.

2. Methodology

The proposed framework integrates three core components: a Contact-Aware Planner, a Trajectory Tracking Controller, and a Specialized Hardware Design.

A. Robot Design and Motion Model

Hardware: A multi-segment, antagonistically driven notched continuum robot with two distinct bending segments (proximal and distal). It is actuated by axial insertion and tendon displacement.
Contact-Aware Motion Model: The authors use a piecewise constant curvature model that solves an energy-minimization problem. It maps commanded joint inputs ( $q$ $q$ ) to robot curvatures ( $\kappa$ $κ$ ) while enforcing:
1. Environmental Constraints: Maintaining a minimum distance between robot points and the environment (point cloud $\Omega$ ).
2. Actuation Constraints: Ensuring the calculated tendon lengths match the commanded lengths.
3. Safety Constraints: Explicitly penalizing contact at the distal tip to prevent vessel perforation.

B. Contact-Aware Planning (Search-Based)

The planner uses a weighted A* search algorithm on a discretized configuration space graph. Key innovations include:

Task-Space Partitions: The workspace is partitioned based on the complexity of motion required to reach the goal (e.g., number of direction changes), rather than just geometric distance. This allows the planner to dynamically adjust the action space (e.g., restricting movements in complex regions) and guide the search.
Cost Function:
- Action Cost: Penalizes joint movement.
- Contact Cost ( $\eta_{contact}$ ): A hierarchical penalty system:
  - Distal Tip Contact: Heavily penalized (prohibited) to prevent injury.
  - End-of-Segment (ECS) Contact: Heavily penalized to prevent kinematic singularities and loss of manipulability.
  - Body Contact: Allowed but minimized to reduce friction and damage.
Heuristic: A task-space heuristic guides the search, incorporating an "approach penalty" to encourage axial alignment with the target vessel at bifurcations, avoiding lateral approaches that risk collision.

C. Control Strategy

To handle model errors and environmental uncertainties, a closed-loop controller is employed:

Inputs: The planner outputs a sequence of target tip poses, joint commands, and Contact-Aware Jacobians.
Jacobian Computation: The Jacobian ( $J_\Delta$ ) is precomputed offline by perturbing joint inputs and measuring tip state changes, accounting for the specific contact configuration at each step.
Feedback Law: A Proportional-Derivative (PD) controller uses the pseudo-inverse of the contact-aware Jacobian to compute feedback corrections ( $u_{FB}$ ), combined with feedforward commands ( $u_{FF}$ ) from the plan. This allows the robot to track the trajectory using only tip pose feedback (minimal sensing).

3. Key Contributions

Differentiated Contact Handling: The first planning approach for continuum robots that explicitly distinguishes between hazardous contact (tip/ECS) and benign contact (body), penalizing the former while permitting the latter.
Task-Space Partitioning: A novel strategy to structure the workspace by motion complexity, enabling dynamic action spaces and improved planning efficiency in constrained environments.
Integrated Planning and Control: A unified system where the planner generates kinematically feasible trajectories and contact-aware Jacobians specifically designed for closed-loop execution, ensuring the plan is executable by the controller.
Hardware Validation: Successful autonomous navigation of a continuum robot through Type I, II, and III aortic arches derived from patient CT scans, achieving 100% success in avoiding dangerous tip contact.

4. Results

The system was validated on 3D-printed anatomical models representing three types of aortic arches (Type I, II, and III).

Success Rate: The robot successfully reached the target in 100% of hardware trials (9 trials total) across all three arch types while avoiding distal tip contact.
Tracking Accuracy: Mean tracking errors were low:
- Type I: $1.9 \pm 0.5$ mm
- Type II: $1.2 \pm 0.1$ mm
- Type III: $1.7 \pm 0.2$ mm
Planning Time: Convergence times ranged from ~5 minutes (Type II) to ~25 minutes (Type I), depending on path complexity.
Ablation Studies:
- ECS Contact Penalty: Removing the penalty for End-of-Segment contact led to kinematic singularities and hardware failure (robot getting stuck). Penalizing it preserved manipulability.
- Body Contact Penalty: Penalizing body contact reduced total navigation time by ~27 seconds and improved tracking accuracy by reducing discrepancies between the theoretical model and physical reality.
- Closed-Loop Control: Open-loop execution failed due to model errors and misalignment; closed-loop control was essential for success.
- Naive vs. Contact-Aware: A naive centerline planner failed in complex environments because it generated kinematically infeasible targets, whereas the contact-aware planner succeeded.

5. Significance and Future Work

Clinical Impact: This work addresses a critical bottleneck in mechanical thrombectomy (stroke treatment). Automating navigation through the aortic arch can reduce procedure time and minimize neurological damage caused by prolonged ischemia.
Safety: By explicitly modeling and penalizing hazardous contact, the framework significantly reduces the risk of vessel perforation, a major complication in endovascular procedures.
Generalizability: The approach is applicable to any soft robot operating in confined, contact-rich environments (e.g., endoscopy, inspection).
Future Directions: The authors plan to extend the 2D formulation to full 3D, reduce planning computation time, and integrate state estimation to infer contact forces for even safer operation.

In summary, this paper presents a robust, safety-critical framework that bridges the gap between theoretical planning and physical execution for continuum robots, enabling reliable autonomous navigation in the most challenging anatomical environments.