A team from EPFL and New York University has built a robotic fish that can be scaled up without being redesigned. The three ScaFi prototypes measure 0.6, 1.1 and 2.9 metres, and essentially one parameter changes between them: the diameter of the rod in the tail. The work was published on 28 September 2026 in npj Robotics.
Key takeaways
- Three prototypes were built, at 0.6 m, 1.1 m and 2.9 m
- Scaling comes down to changing the tail-rod diameter proportionally, while the drive mechanism stays the same
- A single motor moves the tail through two tendons that cross near the base, producing an S-shaped wave
- Swimming kinematics stayed consistent across all three sizes
- Energy efficiency did not scale as predictably as the motion did
One parameter instead of a new design
In robotics, making a design bigger usually means starting over: different gearing, different motors, different stiffness. ScaFi was conceived parametrically. The front is rigid, the tail is a flexible fibreglass rod, and control reduces to tendon drive — two tendons pulled by one motor.
To change length, the designers scale the rod diameter rather than the whole mechanism. That lets the same motion pattern reproduce at every size. The smallest version generated a wake in the water resembling the one a real fish leaves behind.
Three sizes, three environments
| Prototype | Length | Test site |
|---|---|---|
| small | 0.6 m | creeks 15–30 cm deep |
| medium | 1.1 m | Swiss streams |
| large | 2.9 m | Lake Geneva |
The choice of environments is not incidental. A shallow creek demands a small hull, an open lake demands range and mass. One design family covers both, which matters for environmental monitoring, where conditions differ from one survey site to the next.
A negative result stated plainly
The authors do not hide that scaling has a limit. The kinematics transferred cleanly, energy efficiency did not. Larger versions needed more powerful motors, because drag and inertia grow faster than simple proportion would suggest. In practice that constraint sets battery endurance — how long the robot can stay in the field.
The study was funded under Horizon 2020 through Marie Skłodowska-Curie grant 945363.
Why it matters
Parametric robot design is an attempt to bring to hardware what configuration gives you in software. If one design family covers 0.6 to 2.9 metres, the cost of entering a new environment drops from a research project to a change of one dimension. For field robotics, where every site has different conditions, that is the difference between a prototype and a tool.
What's next?
- Non-linear scaling of energy efficiency remains an open problem and caps endurance on the larger builds
- Testing covered streams and Lake Geneva — there is no data on salt water or stronger currents
- The paper describes a design method, not a product: no build costs or deployment plans were given
