A spring between the gearbox and the world

In a conventional electric joint the load is bolted, through a reducer, more or less directly to the motor. That is ideal for position accuracy and terrible for everything a humanoid does with its body: every footstrike, stumble and unplanned contact travels straight back through the gear teeth, and the joint can only guess at the force it is applying by watching motor current through a high gear ratio full of friction.

Gill Pratt and Matthew Williamson’s 1995 MIT paper proposed breaking that rigid chain deliberately: place an elastic element in series between the transmission and the load. The spring does four jobs at once. It filters shock loads before they reach the gearbox. It lowers the output impedance the world feels, so contact is soft even when the controller does nothing. It stores and returns energy across a gait cycle. And, most usefully, it turns force control into position control: measure how far the spring has deflected, multiply by its stiffness, and you have the true output torque, no load cell required. The price, spelled out in the same paper, is bandwidth. A soft element between the motor and the load means the actuator cannot render large forces at high frequency; the spring that protects the gears also low-pass-filters the robot’s strength.

Valkyrie: the most spring-loaded humanoid ever built

The fullest expression of the idea in a humanoid is still NASA’s Valkyrie, the 1.87 m, 129 kg robot NASA Johnson built with the University of Texas for the DARPA Robotics Challenge era. Of its 44 actuated degrees of freedom, 25 are series elastic: rotary SEAs in the shoulders, hips, knees and waist yaw, and linear SEAs working in parallel pairs at the ankles, wrists and waist pitch-roll. Nine joints are conventional rigid drives and twelve are the tendon-driven finger joints of the hands.

The implementation details show how literally the spring is a sensor. Valkyrie’s rotary SEAs put a custom torsion spring behind a harmonic drive and read its twist with Renishaw optical encoders; the linear SEAs run commercial die springs in series with roller-screw drivetrains, with load cells in the output linkage as a second, redundant force measurement. Spring stiffness for each joint was chosen so the spring reaches a fixed target deflection at that joint’s peak torque. The team’s stated reason for the whole architecture: DRC tasks meant movement and manipulation in uncertain environments, where compliant control beats high-gain rigid control, and the spring brings low output impedance, shock tolerance, energy storage and integrated force sensing in one part.

Who puts elasticity where in 2026

Series elastic actuation did not become the industry default. Most commercial humanoids in 2026 run stiff joints and get their force sense from torque sensors or motor current instead. But the spring survives in three distinct forms: inside the actuator (Apollo, Valkyrie’s descendants), as passive elements in the leg linkage (Agility’s Cassie lineage), and, in spirit, as the deliberately low mechanical impedance of quasi-direct-drive joints.

Why Apollo kept the spring

The straightest line from Valkyrie to a robot you can buy runs through Austin. Apptronik’s co-founder and CTO Nick Paine was lead author of the Valkyrie actuator-control paper, and the company’s Apollo is, by its own description, built on force-controlled series elastic actuators. The selection was not sentimental: Paine has said the team built more than 30 unique electric actuators — “liquid cooling, cable driven, series elastic, parallel elastic, quasi-direct drive” — before honing the approach used in Apollo, which ships with just three actuator types, linear units at joints like the elbow and knee, and a target price well under $100,000. Apptronik’s argument for the architecture is the same one NASA made, translated into warehouse economics: a robot that works next to people, and gets shoved by the physical world all day, should be mechanically soft and force-controlled at the actuator level, not just in software.

The counter-argument is the one that won most of the market. The MIT Cheetah’s proprioceptive, quasi-direct-drive approach showed you can get impact mitigation and force sensing without any spring at all, by keeping the gear ratio low enough that the motor itself is back-drivable and its current is an honest torque signal. That design is simpler, shorter and easier to mass-produce, which is why it became the template for the low-cost rotary actuators that dominate Chinese humanoids. Between the two sit the high-ratio harmonic joints of precision-focused robots, which restore force sensing with a dedicated torque sensor rather than a spring. The trade-offs line up like this:

• Shock tolerance. The SEA is the most protective architecture there is for a harmonic drive: footstrike energy compresses a spring instead of shock-loading a flexspline. QDD spreads the hit by having little reflected inertia; rigid high-ratio joints simply take it.
• Force fidelity. A spring read by two encoders measures true output torque continuously and cheaply. Torque sensors do the same for stiff joints at extra cost; motor-current estimates are free but degrade with gearbox friction.
• Bandwidth. This is the SEA’s bill. The spring low-pass-filters force output, so very dynamic motions — the jumps and flips that quasi-direct-drive robots show off — favour stiffer architectures.
• Parts and length. A spring, its guides and a second encoder add mass, axial length and a fatigue-critical component to every joint — real money at 25+ joints per robot, and one reason stiff joints with torque sensors win at scale.
• Energy. Springs store and return gait energy for free. Agility kept exactly this benefit while dodging the actuator complexity, by leaving the springs in Cassie’s and Digit’s leg linkages instead of inside the motors.

FAQ

Sources