Bending to the Task: Smarter Control for Soft Robots

by

Author: Denis Avetisyan

Researchers have developed a new dynamic model that dramatically improves the precision and predictability of soft robot movements.

This work demonstrates a control-affine representation of interconnected elastic rods-originally a physics-based model-allowing for the control of underactuated soft robots, achieved through an explicit mapping of actuator inputs to applied forces, and implemented using a Discrete Elastic Rod (DER) model which approximates the rod’s structure with a system of point masses [latex] [/latex].
This work demonstrates a control-affine representation of interconnected elastic rods-originally a physics-based model-allowing for the control of underactuated soft robots, achieved through an explicit mapping of actuator inputs to applied forces, and implemented using a Discrete Elastic Rod (DER) model which approximates the rod’s structure with a system of point masses [latex] [/latex].

A reformulated Discrete Elastic Rod dynamics model enables accurate trajectory generation for underactuated soft robots with pneumatic actuation.

While soft robots offer compelling advantages for compliant manipulation, accurately modeling their dynamics remains a significant challenge due to their inherent deformability and underactuation. This work, ‘Trajectory Generation for Underactuated Soft Robot Manipulators using Discrete Elastic Rod Dynamics’, introduces a novel control-oriented reformulation of Discrete Elastic Rod (DER) dynamics, enabling the generation of dynamically feasible trajectories that respect actuation constraints. By yielding a control-affine representation grounded in first-principles force-deformation relationships, our approach demonstrably improves trajectory tracking performance over traditional methods-as validated through experiments on a pneumatic soft robotic limb. Could this framework unlock more sophisticated and robust control strategies for a wider range of soft robotic applications?

The Inevitable Complexity of Soft Machines

Conventional robotics typically achieves manipulation and locomotion through a multitude of actuators, each corresponding to a specific degree of freedom. This approach, however, proves problematic when applied to soft robots – machines constructed from highly flexible materials. The very characteristic that defines these robots-their adaptability and ability to conform to complex environments-introduces an infinite number of possible configurations for a given set of actuator commands. Unlike rigid robots where each actuator directly dictates a specific movement, soft robots distribute force throughout their compliant bodies, creating a fundamentally different control problem. This disparity necessitates a shift in robotic control strategies, moving away from precise, direct control towards methods that can harness and predict the complex interplay between actuators, material properties, and resulting deformations.

Underactuated soft robots, distinguished by possessing fewer actuators than degrees of freedom, present a unique challenge in both modeling and control. Unlike traditional robots where each joint has a dedicated motor, these systems rely on the inherent compliance of their materials to achieve complex motions with limited control inputs. This disparity creates an infinite number of possible configurations for a single applied force, making it exceptionally difficult to predict the robot’s behavior. Establishing a definitive relationship between actuator inputs and resulting robot pose requires navigating this vast configuration space, demanding computationally efficient algorithms and innovative control strategies that can effectively harness the benefits of passive compliance while maintaining reliable performance. The inherent ambiguity necessitates robust methods for state estimation and control design, moving beyond techniques applicable to rigidly defined robotic systems.

Predicting how a soft robot will deform under applied force – the force-deformation relationship – is paramount to achieving reliable control, yet presents a substantial computational hurdle. Unlike rigid robots with predictable movements, soft robots possess an infinite number of possible configurations for a single input, making traditional modeling techniques inadequate. Establishing this relationship requires either meticulously detailed finite element analysis, which is extraordinarily time-consuming even for simplified geometries, or extensive data-driven approaches. These data-driven methods demand vast datasets of force and deformation measurements under diverse conditions to avoid inaccuracies and ensure generalization to new, unseen scenarios. Furthermore, the non-linear material properties inherent in many soft robotic materials-such as hyperelastic polymers-exacerbate the complexity, demanding sophisticated constitutive models and increased computational power to accurately capture the material’s behavior under stress. This challenge necessitates the development of novel algorithms and efficient computational strategies to enable real-time control and robust performance of underactuated soft robots.

The system utilizes computer vision to determine the robot's pose and employs a cascade control architecture, where a DER model generates desired pressure setpoints [latex]ar{u}[/latex] that are then precisely tracked by a low-level feedback controller using proportional valves and sensors.
The system utilizes computer vision to determine the robot’s pose and employs a cascade control architecture, where a DER model generates desired pressure setpoints [latex]ar{u}[/latex] that are then precisely tracked by a low-level feedback controller using proportional valves and sensors.

Discrete Elastic Rods: A Pragmatic Compromise

The dynamics model utilizes a Discrete Elastic Rod (DER) approach to represent the soft robot’s structure, offering computational advantages over traditional continuum mechanics methods. This discretization involves segmenting the robot into a series of connected, rigid links, each possessing defined length, mass, and inertial properties. The elasticity of the continuum is then approximated by spring-damper systems connecting these links, allowing for modeling of bending and torsional deformations. By transitioning from partial differential equations – characteristic of continuum mechanics – to a system of ordinary differential equations governing the discrete links, the computational complexity is significantly reduced, enabling faster simulations and real-time control implementations. This method inherently introduces discretization error, but parameter selection – specifically the number of segments and spring constants – allows for tunable accuracy and computational trade-offs.

The Control-Affine Representation extends the Discrete Elastic Rod dynamics model by explicitly incorporating actuation terms into the system equations. This formulation expresses the system’s dynamics as a function of state, time, and control inputs [latex] u [/latex], typically in the form [latex] \dot{x} = f(x, t) + g(x, t)u [/latex], where [latex] x [/latex] represents the system state, [latex] f [/latex] describes the natural system dynamics, and [latex] g [/latex] defines how the control inputs affect the state. This direct incorporation of actuation allows for the computation of control inputs required to achieve desired system behavior without requiring computationally expensive inverse dynamics calculations, facilitating real-time control applications.

Traditional soft robot control often relies on inverse dynamics calculations to determine the necessary joint torques or actuator forces to achieve a desired motion. These calculations, which involve solving for the inverse of the robot’s mass matrix and accounting for Coriolis and centrifugal forces, are computationally expensive and can hinder real-time performance. The presented control-affine dynamics model circumvents this limitation by directly mapping control inputs to the robot’s state derivatives. This direct formulation eliminates the need to explicitly compute inverse dynamics, resulting in a significant reduction in computational load and enabling faster, more responsive control loops essential for the practical implementation of soft robot systems.

The proposed Direct Euler-Radau (DER) formulation generates dynamically feasible trajectories [latex]{\bm{q}(t), \bm{u}(t)}\_{t=0}^{T} [/latex] by directly computing actuation inputs from control-affine dynamics, whereas the Phase-Consistent Control (PCC) baseline relies on a calibrated mapping [latex]{\Lambda}[/latex] to translate virtual torques [latex]{\bm{\tau}(t)}[/latex] into inputs.
The proposed Direct Euler-Radau (DER) formulation generates dynamically feasible trajectories [latex]{\bm{q}(t), \bm{u}(t)}\_{t=0}^{T} [/latex] by directly computing actuation inputs from control-affine dynamics, whereas the Phase-Consistent Control (PCC) baseline relies on a calibrated mapping [latex]{\Lambda}[/latex] to translate virtual torques [latex]{\bm{\tau}(t)}[/latex] into inputs.

Trajectory Generation: A Dance with Underactuation

Trajectory generation utilizes a Control-Affine Representation to address the challenges posed by the robot’s underactuation. This mathematical framework defines the robot’s dynamics as a combination of linear and nonlinear terms, allowing for the decoupling of controllable and uncontrollable degrees of freedom. The system dynamics are expressed as [latex]\dot{x} = f(x) + g(x)u[/latex], where [latex]x[/latex] represents the state vector, [latex]u[/latex] is the control input vector, and [latex]f(x)[/latex] and [latex]g(x)[/latex] define the drift and input matrices respectively. By explicitly modeling the underactuated nature of the robot within this representation, the trajectory generation process can focus on optimizing the controllable inputs to achieve desired motions while respecting the inherent constraints of the system, thereby enabling accurate and feasible trajectory planning.

Trajectory validation utilizes the established robot model to assess ‘Dynamic Consistency’, a critical factor in ensuring predictable system behavior. This process involves simulating the generated trajectory and comparing the resulting state predictions – position, velocity, and acceleration – against the desired trajectory points. Discrepancies exceeding defined tolerances indicate a lack of dynamic consistency, prompting trajectory regeneration or modification. Validating against the model effectively verifies that the planned movements are physically realizable by the robot, accounting for its kinematic and dynamic limitations, and preventing unstable or inaccurate execution.

Comparative analysis against a Probabilistic Collision Checking (PCC) model demonstrates the efficacy of the proposed trajectory generation method. In asynchronous bending scenarios, the approach yields a 60% reduction in tracking error, achieving a mean error of 0.64 cm compared to the PCC model’s 1.59 cm. Across all tested scenarios, the mean tracking error for the proposed method is 0.64 cm (DER), while the PCC model exhibits a mean error of 1.13 cm. These results indicate improved trajectory accuracy and reduced deviation from the desired path when utilizing the proposed control scheme.

Dynamic Equivalent Reduction (DER, red) more closely tracks the reference trajectory (yellow) than Principal Component Capture (PCC, blue), as evidenced by the overlap indicated in orange, demonstrating superior tracking accuracy.
Dynamic Equivalent Reduction (DER, red) more closely tracks the reference trajectory (yellow) than Principal Component Capture (PCC, blue), as evidenced by the overlap indicated in orange, demonstrating superior tracking accuracy.

From Simulation to Reality: The Limits of Control

The developed control algorithms transitioned from simulation to reality through direct implementation on a physical soft robotic platform, leveraging the principles of pneumatic actuation. This approach utilizes pressurized air to inflate and deflate chambers within the robot’s structure, enabling fluid and adaptable movements. By directly controlling air pressure to each chamber, the system achieves precise control over the robot’s bending, twisting, and elongation-characteristics vital for navigating complex environments and manipulating objects. This hardware execution allows for real-time testing and validation of the algorithms, demonstrating their feasibility and robustness in a physical setting and paving the way for practical applications in areas like minimally invasive surgery or delicate object handling.

Accurate estimation of a soft robot’s configuration is critical for effective control, and this is achieved through the integration of computer vision techniques. The system utilizes visual data to track the robot’s posture and deformation in real-time, providing essential feedback for closed-loop control algorithms. This visual feedback allows the robot to understand its own state – position, orientation, and degree of bending – even without direct physical sensors embedded within its flexible structure. By processing images, the system can dynamically adjust actuation parameters to compensate for unpredictable external forces or changes in the environment, enabling precise and robust manipulation tasks. This approach moves beyond open-loop control, allowing the robot to react intelligently and maintain stability during interaction with objects and its surroundings.

Rigorous experimentation confirms the framework’s capacity for dependable manipulation even under challenging conditions. The soft robotic system successfully resisted and compensated for external loading – forces applied to the robot during operation – maintaining its intended trajectory and grip. Furthermore, the control system adeptly managed contact interactions with objects, allowing for nuanced adjustments and stable grasping without compromising precision. These results demonstrate a significant advancement in soft robotics, enabling the robot to perform tasks requiring both strength and delicacy, and paving the way for applications in dynamic and unpredictable environments. The system’s robust performance stems from the integration of computer vision feedback with the pneumatic actuation, allowing it to react and adapt to unforeseen disturbances in real-time.

DER simplifies dynamic modeling by representing a system's state with point masses and edge angles [latex]\mathbf{q}[/latex] and [latex]\bm{\phi}[/latex], which fully define bending angles, material frames, and reference frames.
DER simplifies dynamic modeling by representing a system’s state with point masses and edge angles [latex]\mathbf{q}[/latex] and [latex]\bm{\phi}[/latex], which fully define bending angles, material frames, and reference frames.

The pursuit of increasingly complex control schemes for soft robots, as demonstrated by this work on Discrete Elastic Rod dynamics, invariably invites a certain skepticism. It’s a predictable pattern: elegant theory, promising simulations, then… production. The researchers attempt to refine trajectory generation through more accurate dynamic modeling, moving beyond the simplicity of Piecewise Constant Curvature. One suspects, however, that even this reformulated DER model will eventually succumb to the realities of unpredictable environments and material fatigue. As Henri Poincaré observed, “Mathematics is the art of giving reasons.” But even the most rigorous mathematical model can’t fully account for the chaos that real-world deployment inevitably introduces. Better one accurate, if somewhat limited, model than a sprawling, brittle attempt at perfection.

What’s Next?

The refinement of Discrete Elastic Rod dynamics, as demonstrated, offers a more nuanced model for soft robot trajectory generation. Yet, the elegance of any dynamic model inevitably encounters the brutal realities of fabrication tolerances and unpredictable environmental interactions. The shift from simulation to consistent, repeatable performance in physical systems remains the critical, and often unglamorous, hurdle. This work rightly addresses actuation, but the true test will be scaling these models to robots with a greater degree of underactuation – and the corresponding increase in parameter sensitivity.

Future efforts will likely focus on incorporating learning-based approaches to compensate for model inaccuracies. However, every abstraction dies in production, and these learned corrections will themselves become brittle in the face of novel conditions. The challenge isn’t simply achieving accurate trajectories, but building systems that degrade gracefully when the inevitable disturbances occur.

Ultimately, the field will need to confront the inherent trade-off between model fidelity and computational cost. The pursuit of ever-more-realistic simulations risks exceeding the capabilities of real-time control systems. Perhaps the most fruitful path lies not in perfecting the model, but in embracing the robot’s inherent compliance and developing control strategies that leverage – rather than fight – its flexibility.

Original article: https://arxiv.org/pdf/2603.22604.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-03-25 08:07