Traditional robotics relies on the "rigid-link" paradigm: a finite number of discrete joints connected by stiff beams. In this world, the state of a robot is a simple list of angles. However, soft robots—constructed from compliant elastomers, fabrics, or biological tissues—break this paradigm. As Della Santina and Duriez explain, these systems are continuous, meaning they can deform at any point along their structure. This results in theoretically infinite degrees of freedom (DoF), rendering traditional joint-based algebraic methods insufficient for precise control.
The Complexity of Continuous Deformation
The primary challenge in soft robotics is that the robot’s body is both the actuator and the structure. Unlike rigid robots, where movement is localized at joints, soft robots exhibit global shape changes in response to local forces. Naughton et al. note that this creates highly non-linear physics where bending, twisting, and shearing occur simultaneously.
Furthermore, the materials themselves introduce "messy" temporal dynamics. Hofer et al. and Groen et al. highlight that elastomers and pneumatic actuators exhibit viscoelasticity and hysteresis—behaviors where the robot's current state depends on its past movements. These properties make it nearly impossible to derive exact inverse kinematic models using first principles alone.
Simulating the Squish
To manage this complexity, researchers use a spectrum of modeling approaches that trade off speed for accuracy:
- High-Fidelity FEM: The Finite Element Method (FEM) treats the robot as a mesh of thousands of small, interacting volumes. While accurate for estimating external disturbances, as shown by Cangan et al., these simulations are often too slow for real-time control.
- Reduced-Order Models: Techniques like the Cosserat rod model, used in the Elastica simulator, simplify the robot into a 1D curve that can still twist and stretch. Qin et al. argue that these models are essential for bridging the gap between computational efficiency and physical realism.
- Differentiable Simulation: Tools like ChainQueen, developed by Spielberg et al., allow gradients to flow through the physics engine. This enables "co-design," where the robot’s physical shape and its control logic are optimized simultaneously.
Finding Order in the Chaos
If a robot has infinite DoF, how can a computer possibly control it? The key insight is that while the possible movements are infinite, the useful movements usually live on a much lower-dimensional surface, or manifold.
Spielberg et al. use Variational Autoencoders (VAEs) to compress the infinite-dimensional state into a compact latent representation. Similarly, Alora et al. propose Spectral Submanifold Reduction (SSMR) to identify the "hidden" low-dimensional surfaces where the most important dynamics reside. By focusing only on these essential movements, researchers can apply optimal control techniques that were previously reserved for rigid robots.
In a more radical approach, Nakajima et al. suggest that we shouldn't try to suppress this complexity at all. Instead, they treat the soft body as a "physical reservoir," using its natural, non-linear vibrations to perform computations "for free," effectively turning the robot's own squishiness into its brain.
Proprioception without Sensors
A robot cannot be controlled if it doesn't know where it is. In rigid robotics, we use encoders at every joint. In soft robotics, placing sensors everywhere is physically impossible.
To solve this, Truby et al. developed a "skin" of kirigami-inspired sensors paired with LSTMs to predict 3D configuration. Other researchers, such as Hofer et al., have turned the robot's eyes inward, placing tiny cameras inside pneumatic cavities. By training Convolutional Neural Networks (CNNs) to recognize the patterns of internal fabric deformation, the authors enable the robot to "feel" its own posture without external cameras.
Learning to Act
Because analytical models are so difficult to derive, many researchers turn to data-driven control. However, soft robots are fragile and slow, making the millions of trials required by standard Reinforcement Learning (RL) impractical.
The current frontier involves hybrid strategies. Li et al. demonstrate that pre-training a controller in a "rough" (low-accuracy) simulator can drastically reduce the amount of real-world data needed. Li further pushes this efficiency using Augmented Neural Ordinary Differential Equations (ANODE), which can learn a continuous model of a soft manipulator from as few as 25 data points.
Through these combinations of physics-informed modeling and efficient machine learning, the field is moving toward soft robots that can perform high-precision tasks—like catching a ball or picking berries—while retaining the safety and adaptability of continuous, compliant bodies.