Underwater robotics has long been trapped in a frustrating trade-off between agility and endurance. For years, the industry standard for increasing battery life focused almost exclusively on hydrodynamics—sleeker hulls, reduced drag, and streamlined surfaces. Yet, as autonomous underwater vehicles (AUVs) move toward long-term environmental monitoring, engineers are discovering that the secret to endurance might not lie in how the robot cuts through the water, but in how the motors themselves consume electricity. This shift in perspective is currently playing out in the development of biomimetic systems that stop trying to swim perfectly and start trying to swim intermittently.
The Neural Architecture of a Scaled Zebrafish
To investigate the intersection of biological movement and mechanical efficiency, a joint research team from EPFL, Duke University, and Portugal's IST developed ZBot. This is not a miniature drone, but a massive biological amplification. The team designed ZBot at a scale 200 times larger than a zebrafish larva, resulting in a physical frame 80cm in length and weighing 2.8kg. By precisely replicating the morphological features, segmented body structure, and center-of-mass distribution of the larva, the researchers created a platform capable of mirroring the fluid dynamic interactions of a living organism at a macroscopic level.
ZBot's propulsion is driven by a flexible tail composed of six servo motors that recreate the undulating motion characteristic of fish. The head houses the central controller, which acts as the robot's nervous system, alongside a high-precision camera and a real-time power meter. This setup allows the team to quantify energy consumption instantly during operation. To ensure the platform could adapt to various experimental needs, they integrated an expandable sensor interface, enabling the study of visual-motor processing and vestibular organ functions. This transforms ZBot from a simple robot into a quantitative laboratory for biological swimming strategies.
Control of this movement is handled by a neural computational model consisting of Central Pattern Generators (CPG), a bout gate, and ventral spinal projection neurons (vSPN). The CPG generates the continuous oscillatory signals that form the basic rhythm of the tail. However, these signals do not go directly to the motors. Instead, they pass through the bout gate, which functions as a switching unit. Inside this gate, a leaky integrator continuously accumulates input signals; only when the signal reaches a fixed threshold does the gate release the CPG's rhythm to the motors. This creates a distinct cycle of bout—the active propulsion phase—and glide—the passive phase where the robot drifts on inertia.
To test how this system handles different physical environments, the team manipulated fluid viscosity to simulate the worlds of various aquatic creatures. In standard water (1 cP), ZBot operated in an inertia-dominated regime with a Reynolds number (Re) between 64,000 and 160,000. When moved to a medium-viscosity fluid (213.9 cP, similar to fruit topping syrup), the Reynolds number dropped to between 37.4 and 448.8. In a high-viscosity fluid (457.0 cP, similar to cleansing oil), the Re plummeted to between 1.0 and 87.5, bringing the robot close to a viscosity-dominated regime. While forward distance dropped to 1/30th of the distance achieved in water due to exponential increases in viscous resistance, the robot's turning ability remained remarkably stable. The rotation angle per bout was approximately 60 degrees in water and 45 degrees in high-viscosity fluid, proving that directional control persists even as the fluid dynamics shift from inertia to viscosity.
The Actuator Efficiency Hypothesis
For a long time, the prevailing theory was that intermittent swimming saved energy because the robot reduces drag by keeping its tail straight during the glide phase. However, the ZBot data reveals a more profound electrical truth: the energy savings are driven by the internal efficiency of the actuators. The researchers identified what they call the actuator efficiency hypothesis, which posits that servo motors, much like biological muscles, follow an inverted U-shaped efficiency curve. This means that motors are least efficient when the load is either too low or too high, reaching peak efficiency only at a specific intermediate load.
When a robot employs continuous swimming at low speeds, the motors operate in a state of constant low load. This forces the system to stay in the low-efficiency zone of the U-curve, where the ratio of power consumption to actual thrust is poor. In contrast, ZBot's intermittent strategy utilizes the bout phase to generate short, powerful bursts of propulsion. These bursts temporarily increase the load on the motors, pushing them into the peak of the inverted U-shaped curve where electrical energy is converted to physical motion most efficiently. During the subsequent glide phase, the motors are shut down entirely, eliminating standby power leakage.
This mechanism allows ZBot to selectively utilize only the most efficient operating windows of its hardware. The cost of this efficiency is speed. Because the robot spends a significant portion of its time gliding rather than actively pushing, its maximum speed is limited to approximately 60% of what it would achieve through continuous swimming. The result is a system that is slower but significantly more sustainable, trading raw velocity for a drastic reduction in total energy expenditure.
This discovery shifts the design priority for underwater robotics. To maximize battery life, engineers must look beyond the external shape of the robot and analyze the load-efficiency curves of their chosen actuators. For missions where operational longevity is more critical than speed—such as long-term environmental monitoring—intermittent swimming is not just an option, but a necessity. By calculating the intermediate load zone where the actuator's efficiency peaks and tuning the bout's amplitude and frequency to hit that target, designers can optimize the trade-off between speed and energy. This approach allows for a meaningful extension of deployment time through software control alone, without requiring expensive hardware overhauls.
