What “quasi-direct” actually means

A true direct-drive joint bolts the limb straight to the motor. It is the ideal transmission — zero backlash, zero gear friction, perfect transparency — but electric motors are torque-poor for their mass, so a direct-drive humanoid knee would need an absurdly heavy motor. The classic fix is a high reduction: a harmonic (strain-wave) or multi-stage planetary gearbox at 50:1 to 160:1 turns a small, fast motor into a slow, strong joint. Most industrial arms and many humanoid joints, including the three rotary actuator sizes Tesla showed for Optimus at 20, 110 and 180 N·m, follow that recipe.

Quasi-direct drive splits the difference. A QDD actuator pairs a large-diameter, short-stack brushless motor with a single-stage reduction — typically a planetary set below about 10:1, sometimes a cycloidal or two-stage design reaching toward 20:1 — often packaged inside the motor’s hollow rotor. The motor does most of the work; the gears only help. Because so little transmission sits between rotor and limb, the joint can be back-driven by the outside world, and the controller can estimate joint torque directly from motor current instead of needing a torque sensor or a spring in series. Roboticists call this proprioceptive actuation: the actuator itself is the sensor.

The Mini Cheetah playbook

The design philosophy was worked out on quadrupeds. The MIT Cheetah’s “proprioceptive actuator” papers argued that for legged robots that hit the ground hundreds of times a minute, transmission transparency matters more than gearbox efficiency on a spec sheet: a low ratio lets the leg absorb impacts instead of shattering gear teeth, and lets the controller feel ground contact through the motor windings at high bandwidth. Ben Katz’s Mini Cheetah, shown in 2019, turned that philosophy into a commodity: twelve identical modules, each a hobby-grade outrunner motor with a 6:1 single-stage planetary reduction delivering about 17 N·m peak — and the first backflip by a four-legged robot.

The module was simple enough to copy, and it was copied everywhere. Open-source clones circulate on Instructables and Hackaday, and off-the-shelf QDD modules from suppliers such as CubeMars and Robstride now power a large share of research quadrupeds and humanoid prototypes, at reported prices of roughly $800 to $3,000 per actuator depending on torque rating. When today’s humanoid makers advertise legs that survive falls, torque-controlled joints, or reinforcement-learning policies transferred straight from simulation, they are running the Mini Cheetah playbook at adult scale.

Reflected inertia: the N² tax

The reason the ratio matters so much is that its costs grow with the square. A gearbox multiplies torque by the ratio N, but the inertia of the spinning rotor, seen from the limb, is multiplied by N² — and so, roughly, are gear friction and torque-ripple effects. A 10:1 joint therefore presents about one hundred times less reflected inertia than a 100:1 joint built around the same motor. That is the difference between a foot that touches down softly and one that arrives like a flywheel, and between a joint the controller can model cleanly and one dominated by friction the simulator never saw.

The Berkeley Humanoid, a 0.85-metre, 12-degree-of-freedom research biped presented at ICRA 2025, is the cleanest humanoid statement of this logic: all four of its custom actuator types use the same 9:1 planetary reduction, chosen, in the authors’ words, to “reduce overall friction and minimize the quadratic amplification of cogging torque, friction, and reflected inertia errors.” Its knee delivers 81.1 N·m peak and its ankle just 9.7 N·m — modest numbers, but predictable ones, which is exactly what its learning-based controller needs to cross the gap from simulation to hardware.

Inside a QDD module

Mechanically, a QDD module is one of the simplest actuators in robotics. Electric motor torque grows rapidly with air-gap diameter, so the motor is built wide and thin — a pancake outrunner or frameless torque motor — and the single planetary stage nests inside or directly behind it: the sun gear on the rotor, three or four planets on a carrier that becomes the output, and a fixed ring gear in the housing. The ratio is set by the tooth counts, and a single stage tops out around 10:1, which is exactly where the QDD sweet spot sits. Add an output bearing, an encoder or two and a driver board, and the module is complete. There is no flexspline to fatigue, no torque sensor to pay for, and — compared with a harmonic drive — very little that a hard impact can break.

Who runs low, who runs high

Commercial humanoids rarely publish gear ratios, but the published torque figures and the teardowns sketch the map. Unitree’s joints are in-house, hollow-shaft permanent-magnet units with dual encoders — the G1’s joints top out at 120 N·m and the full-size H1’s M107 knee motor at 360 N·m — and teardowns of the G1 describe compact low-ratio planetary transmissions in the Mini Cheetah tradition rather than harmonic drives. Tesla went the other way for Optimus’s rotary joints, pairing its motors with harmonic-type reducers in three sizes, and put the highest loads on linear roller-screw actuators instead. Research platforms sit at the extremes: Berkeley Humanoid is pure 9:1 QDD, while many manipulation-first robots stay high-ratio throughout.

The pattern is consistent: the more a robot’s pitch depends on dynamic legs — running, hopping, recovering from shoves, falling without breaking — the lower the ratios in its lower body. The more it depends on carrying and precise placement, the more its joints look like industrial ones. The knee of the Unitree G1, whose teardown we covered separately, and the wrist of an assembly-line humanoid are answering different exam questions.

The heat bill

The tax the QDD joint pays is thermal. With only a 6:1 or 9:1 reduction behind it, the motor itself must supply most of the joint torque, and resistive losses in the windings grow with the square of motor current — which is to say, roughly with the square of the torque you are asking for. A high-ratio joint holding a load barely works its motor; a QDD joint holding the same load cooks. That is why QDD spec sheets quote spectacular peak torques next to much smaller continuous ratings, why quasi-direct-drive robots are superb at brief, violent efforts — jumps, sprints, catches, falls — and why a humanoid that must stand all shift holding a payload will not want low-ratio knees.

Designers attack the heat bill from three directions: bigger and wider motors that make torque with less current, active cooling of the stator — the cooling plumbing was one of the notable findings in the Unitree G1 teardown — and duty cycles that let policies exploit short bursts. Researchers are also probing the middle ground: work published in 2025 on backdrivable Wolfrom-type gearboxes aims to keep torque-sensorless control while pushing the ratio well past the QDD range, which would blunt the sharpest trade-off in the table above. Until something like that ships, the choice stays honest, and it stays joint by joint.

When each side wins, in practice:

• QDD wins where contact is the point: impact-heavy locomotion, push recovery and falls, reinforcement-learning policies that must transfer from simulation with low friction and low reflected inertia, and safe physical interaction where the joint must yield and sense through current rather than a force sensor.
• High ratios win where sustained load is the point: holding payloads at standstill, slow precise manipulation, energy efficiency over a full shift, and compact wrist and hand joints where a small motor with a big reduction packages better than a wide pancake motor ever can.

FAQ

Sources