A research team at ETH Zurich has developed a small robot that jumps and walks in liquid, built by designing cultured mouse muscle and a soft artificial scaffold together as a single system. Electrical stimulation makes the muscle contract, which bends the scaffold and produces movement or lets the robot grip objects. The study examined not only the contractile force of the muscle itself, but also how the muscle's placement, the scaffold material, and the way the two are bonded affect performance, using both computation and experiments. However, finding a design that achieves the largest deformation is a separate matter from reliably manufacturing actuators that deliver the same performance.
The results were reported by Aiste Balciunaite, Mike Y. Michelis, and colleagues from the laboratory led by Robert K. Katzschmann. The peer-reviewed paper was published in Nature Communications, volume 17, article number 10732 (DOI: 10.1038/s41467-026-77655-1). The study tests a design approach for "biohybrid robots," which combine living muscle with artificial materials, one that takes into account even the bond between the muscle and the scaffold.
How Muscle and Scaffold Are Bonded Shapes Movement
The team modeled the contracting muscle and the non-contracting scaffold as collections of small three-dimensional elements. They then used an evolutionary algorithm to search for the structure that deforms the most as shape and material placement were varied. In calculations that treated the muscle as a material that contracts like a spring, a two-layer structure of stacked muscle and scaffold was selected as a strong design. When the muscle on one side contracts, the scaffold on the other side resists the contraction, so the whole actuator bends.
That said, adding more muscle does not necessarily produce more bending. In calculations with the stiffness ratio of muscle to scaffold fixed at 1, deformation was greatest when the muscle made up 40–50% of the structure's total height and spread across its entire length. This proportion was obtained under the condition of constant material stiffness and is not a universal optimum applicable to other material combinations.
Where the muscle and scaffold are bonded also proved to be an important factor in deformation. The team varied the bonded area in calculations and confirmed that deformation ranged from 1.46 to 2.44 mm. Figure 3C of the original paper (PDF page 4) compares the following four conditions.
| Muscle–scaffold bonding condition | Calculated deformation |
|---|---|
| Entire surface bonded | 2.44 mm |
| About half of the center detached | 1.70 mm |
| Bonded at both ends only | 1.46 mm |
| About half of the outer edge detached | 2.16 mm |
These are comparisons of different bonded areas in the same two-layer model, not values measured by actually peeling apart cultured tissue. Releasing the bond at the center produced less deformation than releasing it at the outer edges. At least in the computational model, this shows that the location of the bond affects deformation, though it does not guarantee that real cultured tissue would show the same differences. Even so, it suggests that the bonding method is an important design element for transmitting the force produced by the muscle efficiently to the scaffold.
Observations of the physical samples also revealed how difficult it is to bond muscle and scaffold stably. On scaffolds without fine grooves on the surface, the cultured tissue peeled away from the edges and did not integrate along its full length. On scaffolds with grooves, by contrast, muscle fibers aligned along the grooves, and that orientation carried over into the tissue above. However, few cells were present near the boundary with the scaffold, and based on tissue section observations, the authors interpret adhesion as being supported mainly by stable bonding between hydrogels. A strong bond is not created by the muscle fibers themselves adhering tightly to the scaffold.
Jumping, Walking, and Gripping Demonstrated with Cultured Muscle
The actuator the team built is a small structure 15 mm long, 3.5 mm wide, and 1.5 mm high. The cells used were C2C12, a mouse-derived myoblast cell line; this was neither a trial in humans nor an experiment attaching the device to living mouse muscle. Cell-containing material was placed on a scaffold made of water-rich hydrogel, and the cells were allowed to proliferate and differentiate over about two weeks.
The scaffold was fabricated using a light-based 3D printing technique called xolography. The printing itself takes about 60 seconds, but it must be followed by steps such as washing and drying the scaffold and culturing the muscle cells. The scaffold surface has grooves 300 μm wide, 500 μm deep, and 12 mm long. On the third day of culture, the medium was switched to one that induces cell differentiation, and half of the medium was replaced every day thereafter. The fact that a scaffold can be printed quickly should be distinguished from the time needed to complete an actuator with living muscle.
Muscle contraction was driven by electrical stimulation from electrodes placed in the culture medium. According to the paper's methods section, the electric field was 3–5 V/cm, the stimulation frequency was 1–10 Hz, and the duty cycle, the fraction of time the signal is on, was 1%, with stimulation conditions selected according to the motion. When stimulation makes the muscle contract, the scaffold bends; when contraction ends, muscle relaxation and the scaffold's elasticity return it to its original shape. Repeating this contraction and recovery is the basis of the various motions.
In the jumping experiment, a single contraction lifted the robot 5.1 mm from the bottom of the container. The deformation of the actuator itself at that time reached 6.6 mm. This was done in liquid, however, which is not the same condition as a robot jumping through air. In swimming experiments with adjusted buoyancy, the robot rose and then sank at stimulation frequencies of 1 Hz and 2 Hz, while at 3 Hz it reached the liquid surface. These results confirmed how muscle contraction frequency affects movement in liquid.
In the walking experiment, buoyancy was adjusted so the robot could stand on the bottom of a petri dish and move forward. Its speed under spontaneous muscle contraction was reported as 6.2 ± 0.6 mm/s. The ±0.6 is a standard deviation calculated from multiple movement trajectories and does not represent the average speed and variability across many independently manufactured robots. The team also ran experiments in which the robot traveled along a passage with pillars and corners, but the robot did not actively control its direction of travel. Depending on its angle of entry, it collided with pillars and bounced off.
Gripping of objects was also demonstrated. An actuator mounted on a support grasped an object 3.3 mm in diameter and dragged it 1.8 cm. When electrical stimulation was stopped, the actuator opened and released the object. While this shows that contraction of cultured muscle can hold and move an object, it is not a mechanism by which the robot autonomously recognizes and transports a target.
Large Deformation Achieved, but Manufacturing Variability Remains a Challenge
One major challenge revealed by the study is variability in performance from actuator to actuator. When five actuators made with scaffolds of the same dimensions were evaluated, their deformation varied widely, from a minimum of 0.5 mm to a maximum of 6.6 mm. This is not a difference in performance before and after improving the same actuator, but individual variation among separately fabricated actuators. A manufacturing problem that the maximum figure of 6.6 mm alone does not reveal was confirmed even in an evaluation of just five units. In addition, the number of movement trajectories used to calculate walking speed should be distinguished from the number of actuators evaluated for deformation.
According to the authors, growing myoblasts into contractile muscle fibers requires appropriately adjusting not only biochemical culture conditions but also mechanical stimulation and the surface shape of the scaffold. Indeed, in the muscle cultured here, indicators of maturity were not always consistent, and muscle fiber diameters did not reach the values expected for mature muscle. Improving culture conditions and scaffold design may raise performance, but the effect will need to be verified in future work. The computational model also does not fully reproduce muscle maturation or interaction with the surrounding fluid. Even if calculations can select a promising shape, they cannot necessarily predict exactly how each fabricated actuator will move.
Supplying nutrients to the muscle tissue is another challenge for building larger robots. In the cultured tissue here, cells were concentrated in the upper region of about 100 μm. The authors explain that this is because there is a limit to the distance nutrients and oxygen can reach by diffusion. The study also showed an assembly method that joins two matured actuators together, and a method of fabricating a structure with two legs arranged at a right angle as a single piece. However, being able to combine multiple actuators is a separate issue from delivering sufficient nutrients to the interior of thick muscle tissue.
The team has made the computational model and search method used in this work available in a GitHub repository. Calculation examples for single muscle contractions and for varying material stiffness are also provided, offering a starting point for verifying the design method. However, releasing the code does not guarantee that another laboratory can build the physical robot and reproduce the same results. Going forward, in addition to independent replication, what will matter is whether actuator-to-actuator variability can be reduced while maintaining the culture environment, so that the intended deformation can be obtained stably and repeatedly. If these challenges can be overcome, the approach of jointly optimizing muscle and scaffold could open the way to applying the design method to more complex biohybrid robots.