The Watt Tax
The actuator choice that made sense at prototype scale is becoming a serious operational liability at fleet scale.
By Phil D
Every joint in a humanoid robot is a tiny furnace. The motor spins at thousands of RPM. The gearbox steps that down to walking speed. And in that conversion, somewhere between 5% and 30% of the electricity never reaches the joint. It becomes heat. It becomes a tax.
We call this the Watt Tax. It is the single largest energy drain in any electromechanical humanoid, bigger than computing, bigger than sensors, bigger than the thermal management system built to handle it. In Part 1 of this series, we established that locomotion alone consumes roughly 650W in a full-scale bipedal robot during a walking work cycle. In Part 2 we mapped the duty cycle. Now we go inside the joint.
What we found is uncomfortable. The industry has settled on harmonic drives as the default gearbox for precision rotary joints—Tesla, Figure, Apptronik, Boston Dynamics. They are compact, zero-backlash, and well-understood. They are also consistently the least efficient option on the table. The consensus choice is wrong. And the companies that figure this out first will build robots that run longer, cost less to operate, and outcompete everyone else.
Let’s start with the hardware.
The Four Contenders
There are only four practical approaches to reducing motor speed and multiplying torque in a humanoid joint. Each is a tradeoff between efficiency, precision, size, and cost. Here is the state of play as of mid-2026, based on our supply chain teardowns and conversations with actuator engineers at three of the five major humanoid programs.
Harmonic Drive (Strain Wave Gear)
This is the incumbent. A wave generator deforms a flexible spline inside a rigid circular spline, producing high reduction ratios (50:1 to 160:1) in a single stage with effectively zero backlash. The torque density is excellent: 30–39 Nm/kg at the actuator level in production units from Harmonic Drive Systems Inc. (HDSI) and Nidec’s FLEXWAVE line. HDSI claims peak power density of up to 545 Nm/kg for the gear component set alone, but this is peak—not rated—and excludes the motor.
The problem: efficiency runs 75–90%, depending on ratio, speed, load, and lubrication. Under partial load—which is where humanoid actuators spend most of their walking cycle—efficiency drops toward the bottom of that range. The flex spline deforms continuously during operation. That deformation absorbs energy. It becomes heat. It wears the part. The physics is straightforward: the spline must bend to transmit torque, and bending steel takes energy whether or not that energy produces useful work.
The harmonic drive is the IBM mainframe of humanoid robotics: dominant, battle-tested, and quietly becoming obsolete.
The Watt Tax Equation
Every 1% of rotary actuator efficiency loss = 5.12W of continuous waste heat per robot (based on 14 rotary joints at 30W avg each, 75% duty cycle).
Over an 8-hour shift: 0.04 kWh lost per actuator. Over a fleet of 1,000 robots: 560 kWh/day to transmission heat.
That’s not an electricity bill problem. It’s a runtime problem. Every watt of heat is a watt not turning a knee joint.
Cycloidal Drive (RV Reducer)
Nabtesco owns this space. Their RV series uses a two-stage cycloidal design: a planetary input stage followed by cycloidal discs that roll inside a ring of pins. The result is a gearbox with hysteresis loss under 1 arcmin, pure backlash of 0.1–0.2 arcmin, and ratios ranging from 30:1 to over 300:1. The precision is unmatched—this is verbatim from Nabtesco’s own technical literature.
Cycloidal drives handle shock loads better than harmonic drives—the rolling-element contact distributes force across multiple pins simultaneously. Efficiency runs consistently 85–93%, higher than harmonic drives at comparable ratios and notably better at the partial-load conditions that dominate bipedal locomotion. The mechanism behind the efficiency gap is instructive: cycloidal drives use rolling contact, harmonic drives use sliding contact under deformation. Rolling wins. Every time.
The tradeoff: they are larger and heavier for the same torque output. The Nabtesco RV-20E, a common industrial robot joint reducer, weighs roughly 4.9 kg for 245 Nm of rated torque. A comparable harmonic drive unit from HDSI weighs about 2.2 kg. For a 28-actuator humanoid, switching to cycloidal drives across all joints adds roughly 25–35 kg of extra mass. That’s a real penalty. But the efficiency gain offsets it—and then some, as we’ll show.
Planetary Gearbox
The most mechanically efficient option: 95–98% per stage. Simple, robust, well-understood from a century of automotive transmissions. The catch: planetary gearboxes provide low reduction ratios per stage—typically 3:1 to 10:1, with some designs reaching 12:1 to 15:1. To reach the 50:1+ ratios needed for humanoid joints, you need three or four stages stacked together. Each stage adds weight, cost, and backlash. A four-stage planetary might deliver 120:1 reduction but with 3–5 arcmin of lost motion—unacceptable for precision manipulation tasks.
Schaeffler has been the most aggressive in trying to solve this. At CES 2026, they unveiled a planetary gear actuator purpose-built for humanoid robots: two-stage planetary gearbox integrated with motor, encoder, and controller in a single compact package. It won the Hermes Award at Hannover Messe in April 2026—a signal that the industrial automation establishment sees this as real. The same month, Schaeffler signed a strategic partnership with Hexagon Robotics to supply both strain wave and planetary gear actuators for the AEON humanoid platform. The deal covers both a supply agreement for rotary actuators and integration of AEON humanoids into Schaeffler’s global production network.
Schaeffler’s bet: planetary precision can be made good enough. If they’re right, the efficiency advantage is decisive.
Direct Drive / Quasi-Direct Drive
The MIT Cheetah approach: low-ratio gearing (typically 5:1 to 10:1) or no gearing at all, using large-diameter pancake motors to achieve torque through geometry rather than reduction. Maximum backdrivability, maximum transparency, maximum efficiency. The motor operates near its peak efficiency point with minimal transmission loss—90–98% for frameless direct-drive configurations.
The problem is torque density. A direct-drive actuator producing 100 Nm of continuous torque is enormous compared to a geared unit delivering the same output. For a humanoid’s knee joint—which sees peak torques of 150–200 Nm during stair climbing—a direct-drive solution would require a motor roughly the size of the entire thigh. The form factor simply doesn’t work for human-scale joints.
Quasi-direct drive (QDD) with integrated cycloidal reducers at low ratios (8:1 to 15:1) represents a middle ground. Several next-generation humanoid designs we’ve seen—still under NDA—are exploring QDD architectures in the hip and shoulder joints where the torque demands are highest. We expect the first QDD humanoid to ship in 2027.
Three paths to actuating a joint. Harmonic drives deliver zero-backlash precision at the cost of efficiency. Planetary gears offer better efficiency but accumulate backlash across stages. Direct drive eliminates the gearbox entirely but demands impractically large motors for human-scale torque. Source: Machine Narratives Research.
Source: Machine Narratives Research — Actuator Efficiency Model. Efficiency measured at rated torque across 30:1 to 120:1 reduction ratios.
Torque density is the other side of the efficiency equation. Harmonic drives pack the most torque into the least mass — a strong argument for precision joints. But at partial load (where robots spend 70% of walking gait), planetary and direct drive efficiency advantages outweigh torque density gaps. Source: Supplier datasheets, teardown analysis.
The Human Baseline
Here is the uncomfortable truth: the human knee is more efficient than any motor-gearbox combination we can build. But not for the reason most people assume.
Human muscle achieves contraction-coupling efficiencies of roughly 68% at the cellular level—measured in the FDI muscle by Jubrias et al. (2008). That puts it right in the range of a harmonic drive in isolation. But that’s not the whole story. The human leg is a system. The quadriceps stores elastic energy in the patellar tendon during stance phase. The Achilles tendon stores and returns elastic energy with each stride—roughly 35% during running and about 6% during walking (Ker et al., 1987). The broader elastic energy storage system—including the patellar tendon, foot arch, and hip ligaments—contributes additional energy return beyond the Achilles alone. The gastrocnemius-soleus complex operates as a variable-stiffness spring, tuned by neural control to match terrain and gait speed. There is no gearbox. There is no motor. There is no thermal management system bolted on to dissipate waste heat.
The result: human walking carries a net metabolic cost of approximately 0.8–1.0 cal/kg/km for a healthy adult. For a 70 kg human walking at 5 km/h, that translates to roughly 280–350 kcal per hour, or 325W of net metabolic power. Of that, only about 25–30% reaches the joints as actual mechanical work—roughly 80–100W. The rest is heat from muscle inefficiency. Compare to a humanoid: 650W electrical input for locomotion, with perhaps 350–420W reaching the joints as mechanical work through a harmonic-drive transmission chain. The human produces comparable mechanical output from significantly less total input power, using a biological drivetrain that repairs itself.
The gap is not small. It is a factor of roughly 2x in total power input for comparable mechanical output. And almost all of that gap is in the transmission. If human muscles had to push through harmonic drives, walking would burn twice the calories.
We are not suggesting that humanoid robots should use tendons. We are pointing out that a 400-million-year-old evolutionary optimization problem has produced a drivetrain with system-level efficiency that no electromechanical actuator can match. The transmission is the bottleneck. Fix the transmission, and you close the gap.
