Robots Are Limited by Their Muscles, So I Spent Four Months Building a Tiny One

The gap between thinking machines and the physical world

AI is already better than most people at writing, math, and coding. But it still can't fold laundry or hand you a screwdriver. Hans Moravec pointed this out back in 1988:

"It is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility." [1]

That gap is basically why this project exists. The intelligence side is getting solved without me, and the body side is not. In 2023 there were 162 robots per 10,000 factory workers worldwide, more than double what it was seven years earlier (74 in 2016) [2]. That sounds impressive until you look at what the number counts. Robot density counts machines against employees, not against tasks, so it tells you how thinly robots are spread through a workforce and nothing about how much of the work they do. And factories are the easiest place for a robot to survive, structured and fenced off. Homes, hospitals, disaster sites, and the inside of the human body are barely touched.

I think this disconnect between capable minds and incapable bodies is one of the biggest technological limits on quality of life right now. The intelligence is increasingly available. The machinery that would let it act on the world is still missing.

My own reason is less abstract. Robotics is about the closest thing to an epitome of what a mechanical engineer gets to do, and as a rising mechanical-engineering student it made sense to spend my four months on the frontier that manages to be both impressive and actually useful to human quality of life [builder, 2026-08-06]. That is the honest version. The numbers below are the defensible one.

This write-up documents the response to that gap I could build myself. Four months, a home lab, a $1,500–2,000 budget, and an attempt to demonstrate and measure one kind of machine muscle, a 3D-printed, low-voltage stick-slip piezoelectric actuator.

Robots are limited by their muscles

Whatever a robot knows, it can only do what its actuators allow. And actuation is hard in a way computation is not, because it has to move real mass against real forces with real energy, at a size and price the application can afford.

The problem gets much worse as robots get smaller. Electromagnetic motors, the default muscle of big robotics, scale down badly. Shrink one and the available torque collapses faster than the size does, so a centimeter-scale robot built around tiny gearmotors ends up slow, weak, and power-hungry. This is a big part of why the microrobotics field has moved to smart-material actuators instead [5, 7].

You can see the cost of this bottleneck in the field's own numbers. The most complete survey of miniature piezoelectric robots to date covers 342 references. Typical operating voltages in it run 60–200 V, untethered machines are rare exceptions, and its main performance table doesn't contain a single power-consumption entry, because the field mostly doesn't report one [7]. The equivalent review of stick-slip actuators admits the same blind spot. Across 184 references there is no energy-efficiency data anywhere [6]. We are building the muscles before we've figured out how to feed them.

So the argument is short. AI needs bodies, bodies need muscles, and the muscles small bodies need are the ones nobody has finished building, especially at voltages and fabrication methods a normal person can afford. That last part is where a home lab can matter.

A field guide to machine muscles

There are more ways to make things move than most engineers ever use. Figure 1 sketches the five families most relevant to small robots, and Figure 2 places them quantitatively, using the comparison data assembled by Liang et al. (2020) in their study of robotic actuators versus biological muscle [3].

Electromagnetic motors and voice coils are the default choice above centimeter scale. They are mature technology, they run efficiently, and they cost little per watt. But a motor is a speed machine, so to get force you add a gearbox, and the gearbox adds mass, backlash, compliance, and cost. The whole package gets worse as it shrinks. Liang et al. put small electric motors around 0.1 kW/kg specific power, versus 1.6–2 kW/kg for hydraulics [3]. Below about 10 grams of robot, practical electromagnetic options mostly disappear.

Pneumatic artificial muscles of the McKibben type are the most muscle-like option, with about 25% strain at 1.16 MPa stress and efficiency around 49%. Liang et al. note that "the strain and stress of this PAM are slightly higher than those of biological muscle" [3, Table 1]. That holds for stress. On strain it runs the other way against the muscle numbers in the same table, which I give below, so I am quoting the authors' summary rather than endorsing it. The problem is everything attached to them. Compressors, valves, and tubing mean the robot is rarely self-contained. (For calibration, biological muscle itself sits near 40% strain, 0.35 MPa, 40 kJ/m³, and 40% efficiency [3, Table 1]. Every actuator in this section is being measured against that baseline.)

A heated shape-memory alloy (SMA) wire changes crystal phase and pulls with "ultra-high stress to 200 MPa" [3]. Its work density of 10 MJ/m³ is the highest of any row in Liang's table. The catch is about 10% theoretical efficiency, plus thermal time constants that limit it to a few hertz. That combination makes SMA great for latches and one-shot mechanisms and painful for anything that has to move continuously.

Dielectric elastomers are soft capacitor sandwiches that squeeze flat and stretch wide under an electric field, with "giant strains of 380%" and efficiency reported at 60–80% [3]. This is the exciting soft-robotics option. The structural drawback is that the fields require kilovolt drive electronics, which are heavy, expensive, and unforgiving on a small platform.

Piezoelectric ceramics (PZT) deform when charged. The strain is tiny, around 0.1%, but stress, speed, and precision are in a different class from everything above. My own stack actuator makes this concrete. The manufacturer-verified datasheet for the CoreMorrow PSt150/5×5/20H gives 20 µm of stroke (±15%) at 0–150 V from an 18 mm long device, which is 0.11% strain, with 1600 N of blocking force from a 5.1 × 5.1 mm cross-section. That works out to about 61 MPa of blocking stress (Derived: 1600 N ÷ 26 mm²), with 60 N/µm stiffness and a 50 kHz resonance [9]. The blocking force and the stiffness do not square with each other. A 60 N/µm stack pushed through 20 µm of free stroke gives 1200 N, not 1600 N, and I am printing both as I have them rather than quietly choosing the one that flatters the argument. Its stored work per volume is ordinary, about 26 kJ/m³ (Derived: ½·k·x² over device volume, same datasheet numbers), and that figure rides on the stiffness, so it moves if the stiffness does. The unusual part is that it can deliver that work thousands of times per second.

Figure 2. (a) Stress vs. strain for the actuator families in Liang et al. 2020, Table 1 [3]. The PZT-stack point is derived from my stack's datasheet [9]. Both axes are logarithmic. The technologies differ by orders of magnitude, which is why no single actuator wins everywhere. (b) Work per unit volume, same sources.

Why I chose piezoelectric

Every actuator family gives something up. What made piezo worth four months is that it combines properties the other families can't match all at once.

It is solid-state. No gears, coils, magnets, valves, or working fluid, just a block of ceramic and two wires. Nothing wears out except the surfaces you deliberately choose to rub together.

Its resolution sits below a wavelength of light. Displacement is a smooth function of charge, and the stick-slip literature reports step resolutions of 2–40 nm [6]. Piezo scanners are what position atomic-force-microscope tips (general knowledge).

It is fast. My small stack resonates at 50 kHz [9], and ultrasonic piezo motors reach 891 mm/s in linear form [5]. Even at 0.11% strain, a device cycling at just 1 kHz has a strain rate around 110%/s (Derived: strain × frequency), double biological muscle's 50%/s [3].

It delivers force without a transmission. Tens of MPa of actuation stress, orders of magnitude above muscle [3, 9], arrive directly and with zero backlash.

It self-locks at zero power. A friction-coupled piezo positioner holds position when you switch it off, and because the ceramic is a capacitor, holding a voltage draws no static current. For a battery-powered robot that spends most of its life standing still, this property matters a lot [6, 7].

And it scales down well. The physics works from 300-µm MEMS devices up to meter-scale optics mounts, and it's one of the few actuation principles that gets relatively better as everything shrinks [5, 7].

The costs are real too, and I want them on the record from the start: roughly 0.1% strain, so the motion has to be converted somehow, 10–25% hysteresis in the voltage-displacement curve, creep under DC hold, ceramic brittleness, and a traditional 60–200 V drive requirement [5, 6, 7; research_context.md]. One comparative review compresses the list to a single sentence, that "the main challenges for piezoelectric artificial muscles include high stiffness, small strain, and relatively high operating voltage" [4]. Most of this project is an argument with that sentence.

The applications piezo has already won say something about where it wins. Camera autofocus rings, inkjet nozzles, diesel injectors, and scanning-microscope stages are all piezoelectric (familiar commercial examples, general knowledge). At the research frontier the same ceramics fly insect-scale flapping robots, swim robotic fish, and drive endoscopic tools through the body [5, 7]. The pattern is that piezo wins wherever precision, speed, and compactness matter more than raw stroke.

Narrowing the bet to stick-slip actuators (SSPAs)

The central problem of piezo robotics is stroke conversion. You have to turn 20 µm of ceramic expansion into centimeters of travel, and the field has three broad answers: geometric amplification with flexure levers, resonance (ultrasonic motors, fast but demanding to machine and tune), and stepping, where you accumulate unlimited travel out of micrometer-sized bites [5]. A flexure is a hinge with no pin in it, a thin section of solid material that bends where a bearing would otherwise pivot, and levering one trades force for stroke by factors of 5–50×, though the stroke it gives you stays finite.

Stick-slip is the simplest stepping scheme, and Figure 3 is the whole theory. Drive the piezo with a sawtooth. During the slow ramp, static friction drags the slider along with the foot. At the cliff of the sawtooth, the foot snaps back in microseconds and the slider's inertia holds it in place while the foot slides underneath. One small step per cycle, repeated at frequency f, gives speed v = f × step. Two inequalities govern every cycle: the foot has to win during the ramp (µ·N > F_drag) and lose during the flyback (m·a_flyback > µ·N). The entire art of the device lives between those two conditions [first_motion_playbook.md; 6].

Why this scheme instead of the other two? Because every one of its advantages helps a home lab and a mobile robot at the same time.

• Of all the indirect piezo drives surveyed in Zhou 2024, stick-slip has the simplest mechanical structure. A piezo element, a friction block, and a sawtooth [5, replication assessment in paper_analyses/001]. No resonance tuning, no machined stator.
• Travel is unlimited and the steps are sub-micron. Xing et al. measured 55 nm steps at 1 V, 96 nm at 2 V, and 187 nm at 4 V on a three-legged SSPA [8]. The review-level range is 2–40 nm [6].
• It self-locks, as promised above. Friction holds the slider when the drive stops.
• The speed is adequate for locomotion in principle. Research SSPAs reach 12–46 mm/s [6], fast enough to walk a centimeter-scale robot across a desk. Whether a printed 15 V version gets anywhere near that is the open question, and G5 below says what would count.

One more advantage belongs to this specific moment. Stick-slip is where the literature's blind spots pile up on top of each other.

The empty intersection

This part comes straight from reading the field's own review papers against each other.

The first blind spot is that nobody has built a mobile stick-slip robot. Li et al.'s 342-reference survey of miniature piezoelectric robots contains zero demonstrated mobile stick-slip machines in its comprehensive performance table. The two stick-slip entries are a stationary positioner and an incomplete row [7; extraction in paper_analyses/003]. Figure 4 plots that table, and the stick-slip region of the map is empty. The nearest thing to an exception I know of is Adibnazari et al.'s tripedal microrobot [12], which the next blind spot raises for a different reason, and I have not yet confirmed from the paper itself whether it walks and whether it steps by stick-slip. Until I have, the claim I can defend is that Li's table holds no mobile stick-slip entry, not that no such machine exists.

The second is that nobody prints them. In Lin et al.'s 184-reference SSPA review, flexures are wire-EDM'd or CNC-machined metal almost everywhere. 3D printing appears exactly once (Adibnazari et al.'s tripedal microrobot [12]), and the review lists "no 3D-printed SSPAs" as an open gap [6]. Zhou 2024, covering the whole piezo-actuator field in 211 references, doesn't discuss additive manufacturing at all [5; noted in paper_analyses/001]. Where 3D printing does show up in piezo robotics, it's used only for the structural body, with commercial PZT bonded on [7, per paper_analyses/003], which happens to be exactly the architecture a home lab can execute.

The third is that nobody goes low-voltage. Lin 2024 reports SSPA drive voltages of 80–400 V across the review, and lists designing specifically for under 30 V with multilayer stacks as an unexplored direction [6]. Li 2023 agrees from the robot side. 60–200 V is typical, which is a big part of why so few of these machines carry their own batteries [7].

Figure 4. Speed vs. mass for the miniature piezoelectric robots in Li 2023's Table 2 (entries reporting both values; extraction in paper_analyses/003) [7]. Resonant machines (blue) own the fast frontier. Non-resonant steppers (orange) sit on the precise-but-slow shelf. The mobile stick-slip robot doesn't appear because none exists in the survey. The star is where this project aims.

I wish the moment were more cinematic. I was at home, just reading the paper, when I noticed the table had no mobile stick-slip robots in it at all [builder, 2026-08-06]. The rest of this section is the case for filling that hole with a printed, low-voltage machine.

Why should SLA resin printing pair well with SSPAs specifically? Four reasons, each checkable.

1. The tolerances almost meet. Wire EDM holds sub-10 µm tolerances that hobby printers can't match. But a calibrated MSLA printer holds roughly 25–50 µm, and the review-level assessment is that this is viable for proof-of-concept flexures that aren't chasing nanometer resolution [6, assessment in paper_analyses/002]. My own printed-fit testing (sla_tolerance_and_fits_guide.md) lands in the same range.
2. Iteration speed becomes the research method. A flexure geometry costs about $2 of resin and prints overnight, where an EDM'd steel flexure means a quote from a vendor and a wait. When the design space is this open, the lab that iterates ten times faster learns ten times faster.
3. Flexures, foot sockets, preload seats, and alignment features print as one monolithic piece, so there are no assembly tolerances stacking up across five machined components.
4. The unknowns are themselves publishable. Polymer flexures raise questions metal never had to answer, among them viscoelastic creep, their contribution to hysteresis, and above all fatigue life. Nobody reports fatigue data in any of the three reviews [5, 6, 7; gap list in benchmark_table.md]. A home lab that runs a printed flexure to 100,000 cycles and reports what breaks is contributing a number the field doesn't have.

The literature already provides stepping stones. Xing et al.'s three-legged SSPA showed that three triangular waveforms offset by 120° eliminate the backward-motion artifact completely, with 0% backslip at all tested frequencies, 580 µm/s at 900 Hz and 100 V, and 2 kg of vertical load capacity, using 3D-printed casing around commercial stacks [8]. Huang & Sun's 2019 lever-amplified stack, the closest same-class precedent to my hardware, demonstrated steps of 0.875 / 1.75 / 3.33 µm at 10 / 20 / 30 V [10], so meaningful steps survive at the bottom of the voltage range. And Liao et al.'s open-source three-channel drive board (Arduino Mega, DAC0800s, TDA2050 amplifiers, roughly $50–80 in parts [11]) showed the electronics can be commodity-grade. Each of these solves one piece. Nothing I have found assembles the pieces into a printed, low-voltage, mobile stick-slip machine, and that empty intersection is where this project aims.

The goals

The mission is to demonstrate and honestly characterize the core motor of a 3D-printed, low-voltage stick-slip piezoelectric actuator, on a home-lab budget, with every result, file, and failure published open-source.

The constraints go first, because they are the point. Four months (April–August 2026). $1,500–2,000 total budget. Fabrication limited to MSLA/FDM printing and hand assembly. Drive electronics limited to the ~15 V p-p my board actually delivers. (Early project documents said "30 V." That figure turned out to be the supply span, not the signal, and the scope settled the amplitude [11; project CLAUDE.md §2].)

The goals, in order.

• G1. First motion, measured. A working motor at my 5 Hz bench frequency moves a few micrometers per second, which no eye can see, so the readout is a home-built Michelson interferometer (520 nm laser, BPW34 photodiode, $449 oscilloscope) that resolves λ/2 ≈ 260 nm per fringe. Status. Achieved, and it took two tries. First travel on 2026-07-30, 2.51 mm of slider travel in 10 minutes at near-zero preload, which works out to ≈0.84 µm per cycle at 5 Hz (Derived), inside the predicted 0.5–1.0 µm band. Then three weeks in which it would not reproduce (§M10, §M11). On 2026-08-20 the motor ran 3.09 mm in 11 minutes forward and, with the sawtooth reversed, 0.54 mm in 13 minutes back [run log 2026-08-20]. Reversing on command rules out motion that is only drift, tilt, or vibration, so the result counts as confirmed. Two qualifications travel with it. The bench changed in between, since the silicon-nitride ball and the lever leg both went into service in that stretch (§M2, §M9), so August 20 shows motion on a revised machine rather than a repeat of the July run. And both travel figures were read with a ruler against a masking-tape zero, not with the interferometer, because the measurement mirror is not yet on the new slider (§M2, §M8). First motion is demonstrated. It is not yet characterized, and the remaining corroboration runs are listed in §M11. Here is how the first run looked from the chair. I sat at home watching the alumina plate creep away from the edge of a strip of masking tape I'd laid down as the zero mark, motion far too slow to watch in the moment, but over ten patient minutes the gap visibly grew [builder, 2026-08-06].
• G2. The curves. Speed vs. preload from 0 N up in 0.1 N steps. Step size vs. drive voltage, compared against the Huang & Sun ladder [10]. Repeatability spreads. Direction-reversal symmetry. These are the figures a future builder needs.
• G3. Metrology anyone can replicate. The interferometer plus capture pipeline (oscilloscope → Python fringe counter), documented well enough to rebuild from hobby parts and a mid-range scope (mine cost $449 [CLAUDE.md §2], and the optics prices weren't recorded, so I'm not going to make them up). The biggest lesson of this project so far is that you cannot debug what you cannot see.
• G4. Report what nobody reports. Wear and galling on printed friction feet. Creep and fatigue of resin parts under cyclic load. The metrics missing from all three reviews [5, 6, 7].
• G5. The path to a robot. Everything above is single-axis on a rail. The target the design points at, using Xing's topology [8] and my three drive channels [11], is a three-legged mobile platform above 2 mm/s, under 30 V, printable by anyone. That is Phase 2, and every Phase 1 measurement is chosen to feed it.
• G6. Open by default. Data, CAD, protocols, and write-ups released (OSF, GitHub, build logs), negative results included. If the motor had never moved, the metrology and the failure analysis were the publication. That standard doesn't change now that it has moved.

The purpose under the goals is the one this chapter opened with. The gap between what machines can think and what they can do comes down to actuators, and the actuators for the smallest and most numerous machines are still up for grabs. A field this open shouldn't belong only to labs with wire-EDM machines and 200-volt amplifiers. One documented bench in a bedroom can put a real number on a chart where there wasn't one. That's the goal. The rest of the write-up is how it's going.

On to §M2 (the system at a glance)