💻 computer science

Trajectory Generation for Underactuated Soft Robot Manipulators using Discrete Elastic Rod Dynamics

This paper introduces a control-oriented reformulation of Discrete Elastic Rod dynamics for underactuated soft robot manipulators that enables the generation of dynamically feasible trajectories, which are experimentally validated to outperform constant-curvature baselines in tracking accuracy under complex actuation conditions.

Original authors: Beibei Liu, Akua K. Dickson, Ran Jing, Andrew P. Sabelhaus

Published 2026-03-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Beibei Liu, Akua K. Dickson, Ran Jing, Andrew P. Sabelhaus

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 teach a very flexible, noodle-like robot arm how to reach for a cup of coffee. This isn't a stiff robot with joints like a human elbow or knee; it's a soft, squishy robot that bends and twists like a piece of cooked spaghetti.

The problem is that soft robots are incredibly hard to control. If you tell a stiff robot to move its elbow 90 degrees, it does exactly that. But if you tell a soft robot to "bend a little," it might curl up like a pretzel, flop over, or wiggle in unexpected ways because it has infinite ways to deform.

This paper presents a new "instruction manual" (a mathematical model) to help these soft robots move precisely, even when they are being pushed, pulled, or squeezed by the environment.

Here is the breakdown of their solution using simple analogies:

1. The Problem: The "Too Simple" vs. "Too Complicated" Trap

To make a robot move, you need a map (a model) of how it works.

The "Too Simple" Map (PCC): Imagine trying to describe a snake's movement by saying, "It's just a series of straight sticks connected by hinges." This is easy to calculate, but it fails when the snake twists or bends in weird ways. It's like trying to draw a cloud using only straight lines.
The "Too Complicated" Map (PDEs): Imagine trying to describe the snake by tracking every single molecule of its skin. This is incredibly accurate, but the math is so heavy that your computer would freeze before the robot even moved. It's like trying to calculate the weather for every single raindrop in a storm before you can predict if it will rain.

The Goal: The authors wanted a map that is accurate enough to be real, but simple enough to run on a real-time computer.

2. The Solution: The "Discrete Elastic Rod" (DER)

The authors use a model called Discrete Elastic Rod (DER).

The Analogy: Imagine the soft robot arm isn't a continuous noodle, but a string of beads connected by tiny, stretchy springs.
- The beads represent the weight of the robot.
- The springs represent the robot's ability to bend, twist, and stretch.
Why it works: This turns the "infinite" problem of a soft noodle into a "finite" problem of beads and springs. It captures the physics of bending and twisting without needing to solve the impossible math of a continuous noodle.

3. The Big Innovation: Connecting the "Brain" to the "Muscle"

The biggest hurdle with soft robots is underactuation.

The Scenario: Imagine a long, flexible snake. You only have two air pumps (actuators) to control it, but the snake has hundreds of "beads" that need to move.
The Old Way: Previous models treated the air pumps as a vague "curvature" command. It was like telling a pianist, "Play a sad song," without telling them which keys to press. The robot had to guess how to turn that "sadness" into specific muscle movements.
The New Way (This Paper): The authors figured out a direct link. They created a formula that says: "If you pump this much air into Chamber A, it pushes these specific beads in this specific direction."
- They treated the air pressure not as a vague shape, but as a direct force applied to specific beads in their "bead-and-spring" model.
- This allows the computer to calculate exactly what air pressure is needed to make the robot follow a specific path, step-by-step.

4. The "Pretzel" Test (Experimental Results)

To prove their model works, they built a real soft robot arm (a pneumatic limb) and tested it against the old "straight stick" model (PCC).

They gave the robot three challenges:

Simple Bending: Both segments bend the same way.
Opposite Bending: One segment bends left, the other right (like a pretzel).
Chaotic Bending: The segments bend at different times and speeds (Asynchronous).

The Result:

The old model (PCC) was like a driver who only knows how to drive in a straight line. When the road curved or twisted, the robot got lost and missed the target by a wide margin.
The new model (DER) was like a driver who understands the physics of the car. It handled the twists, the turns, and the chaotic movements with much higher precision.
The Stats: In the hardest test (the chaotic "pretzel" movement), the new model was 60% more accurate than the old one.

Summary

Think of this paper as the difference between giving a soft robot a vague wish ("Go there!") versus a detailed GPS route that accounts for the car's suspension, weight, and engine power.

By breaking the soft robot down into a chain of beads and springs, and mathematically linking the air pumps directly to those beads, the authors created a system that can plan complex, dynamic movements for soft robots. This means in the future, soft robots could safely interact with humans, handle delicate objects, or navigate messy environments with the precision of a surgeon, rather than the clumsiness of a floppy noodle.

1. Problem Statement

Soft robots excel in contact-rich and unstructured environments due to their compliance, but this property makes accurate modeling and motion planning difficult. Existing approaches face a trade-off:

High-fidelity models (e.g., Cosserat rods) are physically accurate but computationally expensive (requiring PDE solutions) and often lack a clear mapping from actuator states to generalized forces.
Reduced-order models (e.g., Piecewise Constant Curvature, PCC) are computationally efficient but rely on strong kinematic assumptions that may fail to capture complex deformations and suffer from an "actuation disconnect" (mapping virtual torques to real-world inputs like pressure).

Furthermore, soft robots are typically underactuated (fewer inputs than degrees of freedom). Current Discrete Elastic Rod (DER) models, while bridging the gap between fidelity and efficiency, are primarily simulation-oriented. They lack an explicit, control-affine formulation that maps physical actuator inputs (e.g., pneumatic pressure) directly to the system's generalized forces, making dynamically feasible trajectory generation for underactuated systems challenging.

2. Methodology

The authors propose a control-oriented reformulation of Discrete Elastic Rod (DER) dynamics specifically designed for underactuated soft robots.

A. Control-Affine DER Reformulation

The standard DER dynamics are given by $M \ddot{q} = F_{int}(q) + F_{ext}$ . The authors reformulate the external force term $F_{ext}$ to explicitly incorporate actuation:
$F_{ext} = D \dot{q} + B(q)\Lambda u$
Where:

$u$ : Actuation input vector (e.g., chamber pressures).
$D$ : Diagonal damping matrix.
$\Lambda$ : Scaling matrix mapping inputs to force magnitudes (identified via calibration).
$B(q)$ : A geometry-dependent actuation map encoding force directions.

This structure renders the system control-affine, allowing for the direct computation of control inputs $u$ required to achieve a desired state trajectory.

B. Derivation of the Actuation Map $B(q)$

To construct $B(q)$ without resorting to complex inverse problems, the authors leverage the observation that many pneumatic soft robots exhibit Piecewise Constant Curvature (PCC)-like behavior under bending.

Theorem 1 (Force Localization): Under PCC assumptions, internal elastic bending forces are localized at the segment boundaries (nodes) and are orthogonal to the centerline. Internal nodes experience zero net bending force.
Corollary 2 (Segment-wise Decoupling): Adjacent PCC-like segments can be treated as independent regarding bending-induced forces if their connecting tangents are collinear.
Resulting Structure: Based on these theorems, $B(q)$ is constructed as a sparse, block-structured matrix. Non-zero entries exist only at the boundary nodes of each segment, representing the direction of the force induced by the actuator. This avoids the need for implicit post-hoc mappings.

C. Trajectory Generation Framework

The authors develop a trajectory generation pipeline that synthesizes state-input trajectories $\{q(t), u(t)\}$ to track a task-space reference $r(t)$ :

Inverse Kinematics: Convert the task-space reference $r(t)$ into a reference configuration trajectory $\bar{q}(t)$ .
Direct Input Synthesis: Unlike PCC methods that compute "virtual torques" and then map them to inputs, the proposed DER method uses the control-affine dynamics to directly solve for the actuation input $u(t)$ :
$u(t) = (B(q)\Lambda)^\dagger \left( M \ddot{\bar{q}} - F_{int}(\bar{q}) + K_p(\bar{q}-q) + K_d(\dot{\bar{q}}-\dot{q}) - D\dot{q} \right)$
This ensures the generated trajectory is dynamically feasible and consistent with the underactuated constraints.

3. Key Contributions

Control-Oriented DER Reformulation: A novel derivation that explicitly incorporates actuation into DER dynamics, yielding a control-affine representation suitable for underactuated soft manipulators.
Theoretical Proof of Mechanics Preservation: A proof (Theorem 1 and Corollary 2) demonstrating that the proposed actuation mapping preserves first-principles force-deformation relationships consistent with classical beam mechanics, specifically localizing forces to segment boundaries.
Experimental Validation: A hardware implementation on a 2-segment pneumatic soft robot demonstrating that the proposed method outperforms PCC-based baselines in trajectory tracking, particularly under complex, asynchronous actuation conditions.

4. Experimental Results

The authors validated the framework on a 2-segment, 4-chamber pneumatic soft robot using computer vision for ground truth. Three motion scenarios were tested:

Case 1: Asynchronous Bending (high coupling complexity).
Case 2: Synchronous Bending (Same Direction).
Case 3: Synchronous Bending (Opposite Direction).

Key Findings:

Tracking Accuracy: The DER model consistently achieved lower tracking errors than the PCC baseline. In the most complex scenario (Case 1), the DER mean error was 0.64 cm compared to 1.59 cm for PCC (approx. 60% improvement).
Stability: DER exhibited lower maximum errors and standard deviations, indicating more stable tracking with reduced oscillations.
Robustness: The performance gap widened in scenarios with complex inter-segment coupling (asynchronous bending), suggesting that the DER model's ability to capture deviations from ideal PCC assumptions is critical for accurate control.

5. Significance

This work addresses a critical gap in soft robotics by enabling dynamically feasible trajectory generation for underactuated systems using high-fidelity physics models.

Bridging Simulation and Reality: It transforms a physics-based simulation model (DER) into a control-ready framework, eliminating the need for ad-hoc mappings between virtual torques and physical actuators.
Hardware Compatibility: By explicitly modeling the actuator-to-force relationship, the method generates control sequences that are directly executable on hardware without requiring heavy feedback compensation to correct for model inaccuracies.
Future Potential: The explicit relationship between inputs and generalized forces opens avenues for sensor-free estimation of external forces and interaction, as the model can infer interaction forces based on the deviation between predicted and actual actuation requirements.