How soft robots work, and why medicine cares

Soft robots are built from materials with low elastic moduli (typically elastomers, with Young's moduli comparable to biological tissue), so that deformation of the body itself produces motion and compliance. They have, in principle, very many degrees of freedom, which makes them a natural fit for continuum-style motion.

Soft robots can move in several different ways, although most use fluids, such as air or water, to create movement.

• Fluid-driven actuation: Fluids are pushed into small spaces, called chambers, inside the robot. As these chambers expand, the robot bends, stretches or twists. The shape and structure of the chambers determine how the robot moves. For example, fibres or stronger layers along one side can guide the robot to bend in a particular direction. Air (pneumatic actuation) is lightweight and quick to respond, while water or other liquids (hydraulic actuation) generally provide greater stiffness and more precise control.
• Cable-driven actuation: Cables or thin wires run through or along the robot. Pulling these cables makes the robot bend or change direction, much like pulling the strings of a puppet. This approach can also allow motors to be positioned away from the flexible part of the robot.
• Smart materials: Some materials can change shape when exposed to heat or electricity. For example, shape-memory alloys contract when heated, while electroactive polymers change shape when an electrical voltage is applied. These materials can create movement without the need for fluid-filled chambers.

Each approach has its own advantages and limitations. The choice depends on factors such as the size of the robot, the movement required, the available space and the forces it needs to produce.

Soft robots need ways to sense their shape, movement and interactions with their surroundings. Unlike traditional robots, their bodies can continuously bend and deform, making sensing more challenging. Several approaches can be used:

• Flexible sensors: Small, flexible sensors can be built into the robot to measure stretching, bending, pressure or contact. These sensors deform along with the robot, helping it understand its shape and how it interacts with objects.
• Electrical sensing: Changes in electrical properties, such as resistance or capacitance, can be measured to detect stretching, bending or contact. For example, a flexible material may become more or less electrically resistant as it stretches.
• Optical sensing: Optical fibres or cameras can be used to track changes in the robot's shape and movement. Some optical fibres detect bending through changes in the light travelling through them.
• Model-based sensing: Instead of adding sensors everywhere, engineers can use information already available within the robot. For example, changes in internal fluid pressure or volume can help estimate how much the robot is bending or how strongly it is pushing against something.

Each approach has its own advantages and limitations. Sensors can add complexity and take up valuable space, while indirect measurements often rely on computer models and calibration. Material behaviour, repeated deformation and sensor drift can also affect how accurately a robot senses its surroundings.

Steering a soft robot means working out which inputs, such as pressures, fluid volumes, cable lengths or voltages, will produce the shape or tip position you want. Because the body bends continuously, with no joints to set to an angle, this is harder than for a traditional robot, and the behaviour is often nonlinear and affected by hysteresis and delays. Several approaches are used:

• Analytical models: These simplify the robot into a shape that can be calculated quickly. A common example is piecewise constant curvature, which treats each segment as bending along a circular arc described by a few numbers, such as its curvature, bending direction and length. These models are fast and easy to use, but they ignore effects such as gravity, contact forces and material behaviour, so they become less accurate in real conditions.
• Finite element simulation: The robot is divided into many small elements, and a computer calculates how each one deforms under pressure and external forces. This captures complex material behaviour and contact much more faithfully, but it is computationally heavy, so it is used mainly for design and testing.
• Reduced-order models and learned surrogates: To control a robot in real time, the model must run fast enough to keep up with it. Detailed simulations can be simplified into reduced-order models, or replaced by learned surrogates trained on simulation data, which give similar answers in a fraction of the time.
• Learning-based and AI-driven control: Artificial intelligence (AI) and machine-learning methods can learn how inputs relate to movement from experimental or simulation data. Examples include neural networks and reinforcement learning, where a controller improves through trial and error. They can capture behaviour that is hard to describe with equations, but they need a lot of data, may not work well outside the conditions they were trained in, and are harder to verify for safety, which matters especially in medical use.
• Closed-loop feedback: A model alone drifts over time, because of material wear, temperature, hysteresis and contact with the body. Closed-loop control compares what the robot is doing with what it should be doing, using sensing of pressure, shape or tip position, and corrects the difference continuously.

The best choice depends on the accuracy needed, how quickly the controller must respond, the sensors available and the safety requirements. Many systems combine approaches, for example a fast simplified model for the main movement with feedback to correct errors.

Endoscopy and navigation

Steerable continuum devices for the GI tract, airways and ear canal, with a focus on reducing contact forces and improving reach.

Minimally invasive and micro-surgery

Miniaturised soft segments, sometimes combined with concentric tube robots, for confined anatomy such as the eye.

Assistive and rehabilitation devices

Soft exosuits and gloves with pneumatic actuation and compliant interfaces.

Translation barriers

Sterilisation or single-use design, biocompatible materials, repeatability, sensing for safety, and regulatory pathways (for example medical device classification) shape what reaches the clinic.