Why build a muscle at all

An electric humanoid joint is a rotary motor, a reducer and a bearing stack, sitting right at the joint it drives. Biology does it the other way round. Muscles pull in straight lines across a joint, they go soft when relaxed, and several of them share the work around a shoulder or a knuckle. A muscle-driven robot copies that layout: actuators lie along the bones, tendons cross the joints, and stiffness comes from pulling two muscles against each other.

The appeal is compliance and packaging. A pressurised fibre has no gear teeth to strip in a fall and no backlash. It can be long and thin, so it fits along a forearm where a motor-reducer stack won’t. The price is everything around it – a pump or compressor, valves, a kilovolt power supply or a heat source, depending on the type.

Real muscle sets the bar. Mammalian skeletal muscle contracts 20–40% of its length, delivers around 50 W/kg in typical use with peaks near 284 W/kg, and reaches strain rates of about 500% per second. Those are the reference figures the HASEL research community measures itself against.

How a fluidic muscle pulls

The oldest working design, the McKibben muscle, is a rubber bladder inside a braided sleeve. Pressurise the bladder and it swells. The braid fibres can’t stretch, so they scissor to a steeper angle: the tube gets fatter and, at the same time, shorter. Fix one end to a bone and the other to a tendon and that shortening becomes a pull. Clone Robotics’ Myofiber works on the same principle – a mesh tube that contracts when it’s filled with pressurised fluid.

One quirk matters a lot for control. A fluidic muscle pulls hardest at the start of its stroke and its force drops toward zero as it approaches full contraction – unlike an electric joint, which can hold its rated torque across the whole range. Festo’s DMSP, the best-documented industrial version, is rated to contract up to 25% of its nominal length (20% for the smallest 5 mm size). At 6 bar its theoretical force runs from 140 N for the 5 mm muscle to 6,000 N for the 40 mm one.

Four artificial muscles, side by side

Each technology gets its contraction from a different energy source. Air, water, an electric field or heat. That one choice decides what has to ride along inside the robot.

Stroke: how far each one contracts

Stroke is the number artificial-muscle papers lead with, because it’s where the technologies differ most. A muscle that only shortens 7% needs a long fibre or a lever to swing a finger joint through 90°; one that shortens 30% can sit closer to the joint.

Stroke alone flatters some entries. The nylon figure is a single-fibre lab peak. Clone’s 30% is a company claim. HALVE’s 7% was measured lifting just 42 g. What a humanoid designer needs is stroke, force and speed at the same load, measured the same way – and very few datasets give all three.

Clone’s bet: a whole body on hydraulic fibres

Clone Robotics is the one company trying to build a full humanoid on artificial muscles. Its Protoclone V1 hangs roughly 1,000 Myofibers on 3D-printed polymer bones for more than 200 degrees of freedom, with water circulating through microchannels in the frame for cooling and inertial and pressure sensors built in. The demo unit runs on pneumatics. Clone says later versions move to hydraulics, which matters: air compresses, and that lag shows up in fast balance corrections.

The product version, Clone Alpha, lists 206 bone analogues, 164 points of articulation, 26 degrees of freedom across hand, wrist and elbow, four depth cameras and a compact 500 W pump. Clone opened preorders for a first batch of 279 units. Its stated plan targets a price under $20,000, surgical-tool dexterity in 2026, human-like walking in 2027 and hotel-butler deployments in 2028, and it has reportedly raised $50 million and is opening a hub in Mountain View.

Treat the muscle numbers as claims. None of the Myofiber figures – 30% contraction in under 50 ms, 1 kg of pull per 3 g of fibre – has been published as an independent measurement, and when Maginative covered Alpha, Clone had not yet shown a complete robot working in public. Whether thousands of fluid-filled fibres can be built cheaply and kept leak-free is still untested at production scale.

Electrohydraulic and polymer muscles: still lab hardware

HASEL actuators – hydraulically amplified self-healing electrostatic – are pouches of liquid dielectric between flexible electrodes. Apply a voltage and the electrodes zip together, push the fluid aside and the pouch shortens. The classic versions need 6–7 kV. The HALVE design cuts that to 1,100 V, averages 50.5 W/kg (about the same as typical mammalian muscle), peaks at 132.7 W/kg at 1,300 V and hits strain rates of 971% per second. Its two-channel power supply weighs 15.5 g.

Force is the weak spot. ETH Zurich’s Soft Robotics Lab published a 19-DoF musculoskeletal hand in March 2026 that drives its fingers with Peano-HASEL stacks mounted in the forearm, routed through tendons and a 1:2 pulley. At 5.5 kV the index finger pushes 0.53 N and the thumb 0.26 N, and joints flex about 30° instead of the 90° the design aimed for. The clever part: the hand senses contact from the actuators’ own operating current, with no extra sensors, which let it grip a paper balloon without crushing it.

Twisted-coiled polymer muscles are the cheapest of the lot. Ray Baughman’s group at UT Dallas showed in Science in 2014 that nylon fishing line, twisted until it coils, contracts 49% when heated, reaches 5.3 kW/kg and lifts about 100 times more than a natural muscle of the same size, over millions of cycles. The problem is that it’s a heat engine. Nylon TCPs typically convert only 0.5–2% of their input energy into work, and they have to cool down before the next stroke; bundling fibres for more force adds thermal mass and slows cooling further.

Why the electric joint still wins, for now

Efficiency first. An electric joint goes from battery to drive to motor to reducer. A fluidic muscle adds a pump and valves to that chain, and a thermal muscle throws away nearly all its energy as heat. For a robot that has to work a full shift on one battery, that gap is hard to close.

Control is the second problem. A motor’s torque tracks its current closely. A muscle’s force depends on pressure or voltage and on how far it has already contracted, so the controller has to model a moving target. And there’s supply. Motors, reducers and encoders come from a deep pool of suppliers already scaling for humanoids; nobody makes artificial muscles at humanoid volume yet.

The opening is in hands and faces, where many small, low-force actuators have to pack tightly and gentle contact matters more than raw power. That’s exactly where the ETH hand and most HASEL demos sit today.

What to watch

• Independent force, speed and fatigue data for Clone’s Myofiber, and the first Alpha units actually delivered.
• Low-voltage HASELs moving from grippers and fingers to joints that carry real load.
• Lifetime numbers measured in robot terms – years of daily use – rather than lab cycle counts.
• Hybrid designs: electric legs and torso, muscle-driven hands.

FAQ

Sources