What’s smaller than a grain of salt, smarter than most lab equipment, and can run for months without a recharge? According to Marc Miskin, assistant professor at the University of Pennsylvania, it’s the world’s smallest fully programmable autonomous robot-shrunk down to the scale of bacteria. “We’ve made autonomous robots 10,000 times smaller,” he said, a feat that has forced engineers to rewrite the rules of both robotics and microelectronics.
At dimensions of roughly 0.2 by 0.3 by 0.05 millimeters, these robots inhabit a physical world where fluid dynamics behave very differently. At this scale, water’s viscosity dominates; as Miskin put it, “If you’re small enough, pushing on water is like pushing through tar.” Traditional micromotors with moving parts fail here-propellers stall, hinges snap, and legs become liabilities. The Penn Engineering team abandoned mechanical motion entirely, instead adopting electrokinetic propulsion. Applying a voltage across platinum electrodes, the robots create an electric field that nudges ions in the surrounding liquid. Those ions drag water molecules along, generating a flow that carries the robot forward. This method operates at around 60 nanowatts, requires no moving parts, and is robust enough for months-long operation in a variety of aqueous environments.
Autonomy on this scale requires onboard computation, and here, too, the University of Michigan’s contribution was crucial. David Blaauw’s group, which is known for setting records on sub-millimeter computers, supplied a processor that runs on just 75 nanowatts-100,000 times less power than a smartwatch. To fit in just a few hundred bits of memory, they created a special instruction set that reduces complex operations such as “sense temperature” or “move for N cycles” into single commands. That architecture minimizes program length while allowing behavior to be reprogrammable. The processor, optical receiver, and memory and actuator control circuitry were all fabricated in a 55-nm CMOS process, with high-threshold transistors to suppress leakage currents.
Power comes from microscopic photovoltaic cells covering most of the robot’s surface. These PVs deliver about 1 μW at 0.8 V with 10% efficiency-typical for CMOS-integrated solar cells-and serve dual roles: energy harvesting and optical communication. Programming is done through light pulses, where every robot has an address. Hundreds can share a fluid environment and still be loaded with instructions selectively. A system of passcodes prevents random illuminations from corrupting memory, while a graphical interface automates program generation for non-specialist users.
Currently, the robots’ sensing capability focuses on temperature with 0.3 °C resolution in under 1 mm³, outperforming most comparable volume digital thermometers. During tests, the robots measured temperature gradients in solution, stored information, and reported it back through modulation of their motion – a sort of “wiggle dance” akin to honeybee communication. They also showed a form of taxis-like behavior: they switch propulsion mode to climb toward warmer areas without the need for external control. Such closed-loop adaptation is rare in microrobotics; most systems rely on fixed, preprogrammed routines or external guiding fields.
Fabrication combines lithographic circuit fabrication with low-temperature back-end processes for adding propelling electrodes and releasing the robots from the wafer in one step. The yields exceed 50%, and the process can be qualified in a standard semiconductor foundry to enable scaling. If manufactured in volume, each robot may cost less than a penny, thereby enabling applications such as cell-level diagnostics, microfluidic sensing, and precision manufacturing.
Working in the microscale regime also necessitated the overcoming of difficulties related to fluid mechanics: given that electrokinetic propulsion efficiency is dependent on solution conductivity, the authors determined that a 5 mM hydrogen peroxide medium provides high mobility, as well as tolerance to variations in conductivity. Translation speeds range from 3–5 μm/s and rotational ones from 0.1–0.3°/s; motion is digitally controlled by setting electrode polarities: four-electrode configurations allow 14 unique motion states to be generated, thus enabling tracing of arbitrary planar paths.
While today’s designs are limited to low speeds, low memory and environmental range, the platform integrates propulsion, sensing, and computation at bacterial scale representing a major engineering milestone. Future designs could achieve 100-fold increases in memory with more advanced CMOS nodes, incorporate additional sensors or actuators better suited for in vivo operation. “We saw that Penn Engineering’s propulsion system and our tiny computers were just made for each other.” Miskin calls this “just the first chapter,” according to Miskin, but it’s a chapter that redefines what autonomous machines can be when built at the scale of life itself.
