Eleven years on oil

When DARPA unveiled the first Atlas in July 2013, hydraulics were not a choice so much as a necessity. The robot had 28 hydraulically actuated degrees of freedom and was tethered to an off-board power supply, because nothing electric at the time could deliver the force a 1.5-metre biped needs at its hips and knees. The 2016 next-generation Atlas cut the tether — an 80 kg machine carrying its own battery, running an electric pump that pressurised the oil driving every joint. The final HD Atlas, at roughly 85 kg, is the robot the world knows from the backflip, parkour and dance videos: proof that a hydraulic humanoid could deliver explosive, whole-body athleticism no electric machine had matched.

And then it was gone. Boston Dynamics retired the hydraulic Atlas on 16 April 2024 with a farewell video that lingered, pointedly, on years of oil leaks spraying from burst lines. The next day the company announced an all-electric Atlas, describing it as stronger and with a broader range of motion than any previous generation. The new machine is built around custom electric actuators using planetary roller screws and high-density neodymium magnets, with Hyundai Mobis lined up to supply actuators at mass production and deployment in Hyundai’s automobile assembly planned for 2028.

The company had run this play before. Its early legged machines — BigDog and its successors — were hydraulic; the robot it actually turned into a product, Spot, is electric. Atlas followed the same commercial logic: hydraulics won the research decade, and lost the product decision.

Two ways to power a knee

The two architectures differ most in what sits between the battery and the joint. A hydraulic humanoid converts electrical energy into fluid pressure once, centrally, then distributes that pressure through lines and servo valves to compact cylinders at each joint. An electric humanoid skips the fluid: each joint has its own motor and transmission, fed by wires, and draws current only when it is actually working.

The case against the pump

On a battery-powered biped, the central pump is a tax on everything. A hydraulic system must hold line pressure whether the robot is sprinting or standing in a queue, and the losses stack up at the pump, in the lines and across every throttling valve. Published engineering comparisons put typical hydraulic system efficiency at 40 to 55 percent, against 75 to 80 percent for electric actuators that draw current only when the joint actually works. On a machine whose entire energy budget hangs from its back, that difference is runtime.

The rest of the indictment is operational rather than thermodynamic.

• Maintenance and leaks. Oil at 1,800 to 3,000 psi finds every weak fitting. Hydraulics need filters, fluid changes and specialised service; electric joints need almost none of it. Boston Dynamics’ own farewell video made the leaks a running joke.
• Thermal ceiling. Standard hydraulic fluid degrades above roughly 82 °C, capping sustained hard running; electric actuators manage heat at each joint with a wider operating window.
• Cost and complexity. A hydraulic humanoid needs custom machined manifolds, servo valves and lines that exist nowhere else in industry, which is a large part of why analyses of the switch describe hydraulics as complex and costly to productise. Electric joints ride a supply chain shared with drones, e-mobility and industrial servos.
• Noise and cleanliness. A pump whines even at rest, and factories, warehouses and homes do not want atomised oil. The electric Atlas is notably quieter than its predecessor — a selling point Boston Dynamics highlights for commercial settings.

What electric had to prove first

None of those complaints are new — they were just worth paying for as long as electric motors could not deliver the force. That gap has been closing for thirty years. Before the 1990s, hydraulic actuation held roughly a hundredfold power-density advantage over electric motors; advances in neodymium permanent magnets and motor design have cut it to something like a factor of ten. What closed the remaining distance in humanoids was the transmission: quasi-direct-drive layouts pair high-torque motors with low gear ratios, and planetary roller screws convert a small motor’s speed into thousands of newtons of push. The electric Atlas leans on exactly these — roller screws and high-density magnets — and Tesla’s Optimus made the roller-screw linear actuator the industry’s most-watched component.

The market has voted. Every humanoid currently shipping or scaling — Optimus, Figure, Unitree’s G1 and H1, Apptronik’s Apollo, Agility’s Digit, Fourier’s GR series, and now the electric Atlas — is fully electric in its limbs. Schaeffler, a century-old supplier whose business was built partly on hydraulic components, frames its humanoid strategy under the title “Goodbye hydraulics.” In 2026 there is no hydraulic humanoid on any commercial price list.

The holdouts: hydraulics at the fingertip

Hydraulics did not die; they retreated to the one place electric motors still struggle — very high force in a very small space. A humanoid finger joint has millimetres to work with, and at that scale fluid pressure beats copper windings on force density by a wide margin.

Sanctuary AI is the clearest example. Its Phoenix humanoid is electric in the body but hydraulic in the hands: 21 degrees of freedom driven by miniaturised hydraulic valves running food-safe oil, which the company says have been tested past two billion cycles without leakage or degradation. Sanctuary’s argument is precisely the old Atlas argument at a smaller scale — superior power density where strength and human-like dexterity have to coexist, for work such as electronics assembly. Clone Robotics goes further in the other direction: its Torso 2 android is actuated by around 910 water-driven artificial-muscle fibres across some 164 degrees of freedom, using water instead of oil specifically to make fluid power tolerable in a household. And in the research world, integrated “smart” hydraulic actuators with built-in electronics are still published as the route to force-control bandwidth that geared electric joints cannot reach.

What it means for buyers and suppliers

For anyone specifying a humanoid in 2026, actuation type is no longer a fork in the road — you are buying electric, and the real questions are one layer down: rotary or linear architecture, reducer type, thermal ratings and who supplies the screw. For component makers the lesson cuts both ways. The hydraulic ecosystem that fed Atlas-era research has no humanoid customer left at the limb scale, but the micro-hydraulic niche in hands is genuinely open, with valve miniaturisation as its gating technology. And the companies that supplied hydraulics know it: Schaeffler’s own literature now walks its customers from cylinders to electromechanical actuators, and Boston Dynamics’ actuator supply is being industrialised by an automotive Tier-1 rather than a fluid-power house.

FAQ

Sources