News

Tendon-Driven Hands vs In-Finger Actuators: The Open Question in Humanoid Hardware

Humanoid hand design is splitting into two serious hardware strategies: remote tendon-driven hands and compact in-finger actuator hands. This guide explains the tradeoffs, failu...

The humanoid robotics industry is not just arguing about legs, batteries, AI models, or factory cost. One of the hardest unresolved questions is at the end of the arm: how should a robot hand actually be built?

By 2026, the serious humanoid builders are splitting into two broad camps. One camp is trying to imitate biology more closely, using motors in the forearm and cable-like tendons to pull the fingers. The other camp is putting more of the actuation inside the hand itself, with motors and transmissions living in the palm, finger base, or phalanges they control.

Both approaches can produce impressive demos. Both can be made to pick up tools, fold fabric, hold cups, sort objects, and manipulate small parts. Neither has clearly won. The important point for buyers, investors, engineers, and robotics researchers is that these two architectures fail in different ways. A hand that looks fluid in a short video may be difficult to manufacture, recalibrate, repair, cool, seal, or run for thousands of hours.

The comparison graphic above frames the split clearly: tendon-driven systems are associated with Tesla Optimus, 1X NEO, and some Agibot hand designs or industry interpretations; in-hand or in-finger actuator systems are associated with Figure, Unitree, Sharpa, and ROBOTIS. The exact public evidence varies by company. Some firms publish detailed hand specs. Others show capability videos while keeping the actuator layout private. Where the source record is not public, this article treats the classification as a working industry read, not a verified teardown.

Why the hand is the bottleneck

A humanoid robot can only be useful if it can touch the human world without special fixtures. Doors, mugs, shirt buttons, dishwashers, storage bins, charging cables, zippers, tools, and cardboard boxes were not designed for robot grippers. They were designed around human hands. That is why so many humanoid companies keep returning to a five-finger hand instead of a simple two-jaw industrial gripper.

The problem is density. A human-sized hand has very little room for motors, gearboxes, bearings, encoders, wiring, tactile sensors, protective skins, seals, thermal paths, and structural members. It must also survive contact. A mobile robot hand gets hit, bent, twisted, pushed into drawers, trapped under objects, and loaded from odd angles. It has to move gently around people while still applying enough force to use tools and carry useful payloads.

That makes the hand one of the most unforgiving pieces of humanoid hardware. The engineering tradeoffs are not cosmetic. They decide whether the robot can learn from human motion, whether it can feel slip, whether it can work near water or food, whether a broken finger can be replaced in minutes, and whether the product can be manufactured in volume instead of built by a lab team.

The tendon-driven approach

A tendon-driven robotic hand moves the heavy actuation away from the fingers. Motors sit in the forearm, wrist, or another remote module. Cables, belts, synthetic tendons, or tendon-sheath routes carry force into the hand. Pull one tendon and a joint flexes. Pull another and the finger extends or abducts. Add differential mechanisms and one motor can influence multiple joints.

This is mechanically familiar because human hands work in a related way. Much of the strength that moves our fingers comes from muscles in the forearm. Tendons cross the wrist and run into the hand. Robotics does not copy biology perfectly, but the core idea is attractive: put mass where there is more room, keep the fingers light, and let the distal hand remain small enough to fit into human spaces.

Why tendon-driven hands are attractive

The first advantage is low finger inertia. If the motors are not inside the fingers, the fingers can be lighter. Lighter fingers are safer in accidental contact, quicker to reverse, less punishing on small joints, and easier to move with delicate force. This matters for home robots, factory robots working around people, and any robot expected to handle fragile objects.

The second advantage is packaging. The palm and fingers can be used for structure, tactile sensors, protective skins, wiring, and compliant mechanisms instead of being dominated by motors and gearboxes. A slim finger can reach into a cabinet, a bag, a cup handle, or a tool grip more naturally than a bulky powered finger.

The third advantage is heat. Motors generate heat. Putting more motor mass in the forearm gives the designer more surface area, more thermal mass, and a better path to manage continuous use. In a finger, there is little room for heat sinking, airflow, or insulation from tactile materials.

The fourth advantage is force transparency. If a tendon transmission is built with low friction and low gear reduction, contact force can travel back through the transmission. That can make the hand feel the world through its own joints instead of relying only on fingertip pads or cameras. This is one reason 1X emphasizes its NEO hand as a read-write interface, not just a position-controlled end effector.

The hard parts of tendon systems

The difficulty is that tendons are not magic wires. They stretch, creep, wear, bend around guides, and change behavior as the wrist moves. Cable friction creates hysteresis: the force delivered at the fingertip is not always the same force commanded by the motor. The same motor command can produce a slightly different joint result depending on direction, load, cable routing, temperature, and wear.

Crosstalk is another serious issue. Tendons often pass through the wrist. When the wrist bends, the tendon path length can change. That can unintentionally flex a finger or reduce available force. Engineers can fight this with careful routing, pulleys, curved guides, pretensioning systems, and compensation models, but those fixes add parts and assembly steps.

Calibration is also harder. A tendon-driven hand needs correct tendon tension, known routing friction, encoder offsets, force models, and often a map from motor position to joint position. If a cable is replaced, stretched, contaminated, or routed slightly differently, the hand may need recalibration. This is one reason serviceability becomes a major commercial question. The smoothest demo hand in the world is less attractive if every repair requires a trained technician to open the wrist and retension multiple channels.

Manufacturing is the final challenge. Tendon systems can look elegant in a prototype and become painful in production. Cable length, crimp quality, pulley alignment, sheath routing, surface finish, guide wear, and assembly order all matter. A factory needs repeatability. A lab can tune one hand. A product company has to make thousands of hands that behave the same way.

Companies using or moving toward tendon-driven hands

Tesla Optimus V3

Tesla has not published a complete final Optimus hand datasheet, but patent reporting in 2026 points to a tendon/cable-driven architecture for the next Optimus hand. The reported design moves actuators into the forearm, routes tendons through a wrist transition, and gives each finger multiple degrees of freedom. The strategic logic is clear: Tesla wants human-like dexterity without building a heavy, oversized hand.

The risk is equally clear. Tesla is targeting high-volume manufacturing and useful factory work. That means the hand cannot be a delicate research mechanism. It has to survive repetitive tasks, dust, impact, tool use, and fast assembly-line motion. Tesla's hardest problem may not be making one hand look human. It may be building a hand that can be manufactured, repaired, and kept in calibration at automotive scale.

1X NEO

1X is one of the clearest public examples of the tendon-driven camp. Its 2026 NEO hand announcement describes 25 degrees of freedom, with 22 fully actuated degrees of freedom in the fingers and palm plus 3 at the wrist. 1X says the system uses quasi-direct-drive tendons through its 1X Tendon Drive with low gear ratios, force control, and backdrivability.

That is an important technical claim because it links dexterity with sensing. 1X is not merely trying to pose fingers. It wants the hand to ask the world questions through force and read the answer back through the same joints. If that survives real home use, it is a strong argument for tendon-driven hands. The open questions are durability, cost, cleaning, field service, privacy in teleoperated homes, and whether the company can keep the hand consistent across large production volumes.

Agibot OmniHand and OmniHand Pro

Agibot is a useful example because public classification is not completely clean. Some third-party summaries group Agibot's OmniHand Pro with tendon-driven or remote-transmission hands. Agibot's own store page, however, describes the OmniHand Pro 2025 as using linkage transmission and a motor plus screw-rod drive method, with 19 total DoF, 12 active DoF, 20 N typical fingertip force, CAN FD communication, and tactile sensing.

For a buyer, that distinction matters. A linkage or screw-driven anthropomorphic hand can still share some goals with tendon systems: human scale, compact packaging, multi-finger control, tactile feedback, and safe interaction. But it should not be treated as proven tendon architecture unless Agibot publishes that mechanism or a teardown confirms it. The broader point is that the market is already producing hybrid designs that do not fit neatly into simple labels.

The in-finger or in-hand actuator approach

The alternative is to move more actuation into the hand itself. In this architecture, motors and transmissions sit in the palm, finger base, or the individual phalanges they move. The exact implementation varies. Some use direct-drive or quasi-direct-drive joints. Some use tiny geared motors. Some use compact screw drives, linkages, or modular actuator cartridges.

The appeal is straightforward: put the actuator close to the joint and the control problem gets cleaner. There are fewer long tendons to stretch, fewer wrist routing problems, and fewer hidden friction losses. A finger module can be designed, tested, replaced, and characterized as a relatively self-contained part.

Why in-finger actuation is attractive

The first advantage is directness. A motor near the joint gives the controller a clearer relationship between command and motion. The system can still have backlash, elasticity, friction, and compliance, but the path is shorter and easier to model than a tendon running through a moving wrist.

The second advantage is serviceability. In a well-designed modular hand, a damaged finger can be swapped without dismantling the forearm. That matters in commercial deployments. Robot hands will break. They will be dropped, jammed, overloaded, pinched, contaminated, and abused by the world. If maintenance takes ten minutes instead of two hours, fleet economics change.

The third advantage is manufacturing simplicity. A direct actuator module can be tested before final assembly. Factories like modules with repeatable electrical tests, known torque curves, known encoder behavior, and limited manual tuning. If every finger module behaves similarly, the hand can scale more easily.

The fourth advantage is software. A local actuator with local sensing is easier to represent in simulation and easier to calibrate in firmware. This can make teleoperation, reinforcement learning, and imitation learning pipelines more predictable.

The hard parts of in-finger actuators

The penalty is mass. Put motors and gearboxes in the fingers and the fingers become heavier. Heavy fingers need stronger upstream joints, consume more energy during fast motion, create higher impact loads, and can make the hand less safe around people. A heavy finger also makes delicate manipulation harder because the mechanism itself contributes more inertia than the object being manipulated.

Thermal design is another constraint. Small actuators working inside small fingers have little room to shed heat. A hand may perform well in a thirty-second demo but derate during continuous sorting, cleaning, or tool use. Heat can also affect tactile sensors, adhesives, seals, lubricants, and plastic components.

Packaging is unforgiving. A human-scale finger is narrow. Once a designer adds a motor, reduction, encoder, wiring, bearing supports, structural shell, fingertip skin, and tactile sensor, there may be little room left for compliance. The finger can become strong but bulky, precise but fragile, or serviceable but less human-shaped.

Impact tolerance also matters. A tendon-driven finger can sometimes give way because the heavy motor is remote and the distal mechanism is lighter. A finger with built-in actuators may expose more expensive parts to direct impact. Good mechanical stops, compliant shells, clutching, current limits, and sacrificial covers become essential.

Companies using or associated with in-hand actuation

Figure 03

Figure's public Figure 03 announcement focuses less on actuator placement and more on what the hand is supposed to enable. Figure says the robot has a redesigned hand system built for Helix, its vision-language-action AI. The company also highlights softer adaptive fingertips, palm cameras for close-range visual feedback, and internally developed tactile sensing designed for durability and high-fidelity contact data.

The supplied industry graphic places Figure in the in-finger actuator camp. Figure has not published a full actuator map in the public source we reviewed, so the safer reading is this: Figure is clearly investing in a heavily redesigned, sensor-rich humanoid hand, but buyers should ask for actuator layout, finger module replacement procedure, duty-cycle thermal data, and tactile sensor durability before treating the design as fully understood.

Unitree Dex5-1

Unitree's Dex5-1 is one of the most concrete public examples in this category. Unitree lists 20 degrees of freedom, 16 active plus 4 passive, with a thumb at 4 DoF and the other fingers at 3 DoF each. It also lists 94 tactile sensors per hand for the Dex5-1P version, a 1,100 g hand weight, USB 2.0 communication, 24 V to 60 V working voltage, and a 1,000 Hz communication rate.

Unitree describes self-developed micro force-controlled composite transmission joints and micro force-controlled joint gear transmission. That puts the Dex5-1 squarely in the compact, modular, in-hand actuation conversation. The engineering question is whether Unitree can combine low cost, robust tactile data, replaceable fingers, and enough thermal headroom for long work sessions.

Sharpa Wave

Sharpa Wave is positioned as a high-end dexterous hand for physical AI research and robot integration. Sharpa lists 22 active degrees of freedom, 1:1 human scale, 20 N fingertip force, more than 1,000 tactile pixels per fingertip, 0.02 N force sensitivity, ROS compatibility, Isaac Sim and MuJoCo support, and durability tests including 2.5 million press cycles and 3,200 impact cycles at 30 g.

The public page does not expose every internal actuation detail, but the product direction is clear: a dense, sensor-rich, developer-facing hand intended to generate manipulation data and run contact-rich tasks. For teams comparing it against tendon-driven hands, the questions are price, replacement modules, thermal behavior, tactile sensor repair, API quality, and whether the hand's high specification translates into reliable task success outside demos.

ROBOTIS HX5-D20

ROBOTIS is unusually explicit. The HX5-D20 is a 20-DoF five-finger hand, and ROBOTIS states that each finger joint is powered by an XM335-T323-T DYNAMIXEL actuator in a direct-drive configuration. ROBOTIS also emphasizes tactile sensing at the fingertips and compatibility with its broader software and actuator ecosystem.

This is the clean argument for the in-finger/direct-drive side. The mechanism is easier to reason about, easier to document, and easier for developers to integrate. The tradeoff is that direct actuator placement has to prove itself on mass, heat, impact resistance, and human-scale packaging.

What the demos do not prove

Most robot-hand demos are selected demonstrations. They show a robot picking up a grape, pouring tea, folding fabric, holding a tool, catching a ball, or turning an object. Those clips are useful, but they do not answer the commercial question by themselves.

A serious hand comparison needs repeated task data. How many successful picks out of 1,000? How many failures after dust exposure? How many hours before recalibration? What happens after the hand is struck sideways? Can a technician replace a finger in the field? Does the fingertip temperature rise during repetitive gripping? How much does tactile performance drift after cleaning? Can the hand still work when the wrist is bent and the arm is moving?

This is where tendon-driven and in-finger systems need different tests. A tendon-driven hand should be tested for cable stretch, friction drift, crosstalk through wrist motion, retensioning time, and behavior after cable replacement. An in-finger actuator hand should be tested for thermal derating, impact survival, module replacement time, gear wear, backlash growth, and power density under continuous use.

The buyer checklist

If a humanoid robot company claims human-level hands, ask for the following before believing the headline:

Actuator location: forearm, wrist, palm, finger base, phalange, or a hybrid layout.
DoF breakdown: total DoF, active DoF, passive DoF, wrist DoF, thumb DoF, and whether fingers are independently controlled.
Service procedure: time to replace one finger, one tendon, one fingertip skin, and one tactile sensor.
Calibration requirement: whether a field repair needs software calibration, manual tensioning, or factory tools.
Thermal limits: continuous gripping time, maximum fingertip temperature, derating behavior, and duty-cycle assumptions.
Tactile data: sensor type, sampling rate, force range, shear sensing, slip detection, and API access.
Durability: press cycles, abrasion tests, impact tests, water or dust rating, cable life, gearbox life, and skin replacement interval.
Task benchmarks: repeated real-world tasks with success rate, not just edited video clips.

What would decide the winner?

The winner will not be chosen by the most impressive hand photo. It will be chosen by economics and task success. A hand wins when it can do useful work, survive daily contact, be manufactured consistently, and be repaired without destroying fleet uptime.

Tendon-driven hands may win if they deliver better dexterity, safer contact, lower finger inertia, and richer force transparency without becoming a service nightmare. In-finger actuator hands may win if they deliver enough dexterity with simpler assembly, easier maintenance, clearer control, and predictable modular replacement.

There may not be one winner. Home humanoids may prefer lightweight tendon-driven hands because safety, softness, and fine manipulation matter more than abuse resistance. Factory humanoids may prefer modular in-hand actuators because field service and uptime matter more than perfect human-like packaging. Research platforms may support both because different learning algorithms need different mechanical assumptions.

The practical conclusion for 2026

The open question of humanoid hardware in 2026 is not whether robot hands can look human. They can. It is whether they can become reliable industrial and consumer products.

Tendon-driven hands are elegant, biologically persuasive, and potentially safer at the fingertip. They also bring hard problems in friction, routing, tension, crosstalk, calibration, and repair. In-finger actuator hands are more modular, more direct, and often easier to control. They also fight mass, heat, packaging, impact loads, and the cost of miniature precision actuators.

That is why this split matters. The hand architecture tells you what kind of company you are looking at. A tendon-driven hand suggests deep vertical integration, manufacturing discipline, and a willingness to solve painful mechanical detail. An in-finger actuator hand suggests modularity, serviceability, and a stronger component ecosystem. Both strategies can work. Both can fail.

For now, the best answer is not to pick a camp too early. The best answer is to demand better benchmarks: repeated household tasks, tool use, cable plugging, dish loading, fabric folding, door opening, failure recovery, field repair time, and thermal endurance. The first humanoid hand architecture to pass those tests at scale will define the next phase of robotics.

Frequently asked questions

Are tendon-driven robot hands more dexterous?

They can be, especially when low finger mass, compliance, and force transparency are important. But dexterity depends on the full system: DoF, actuation bandwidth, tactile sensing, control software, calibration, and durability. A poor tendon hand can be worse than a good direct actuator hand.

Are in-finger actuators easier to repair?

Usually, yes, if the product is designed with replaceable finger modules. The advantage disappears if the hand is tightly packed, glued, sealed, or requires factory calibration after every repair.

Which approach is better for home humanoid robots?

Home robots need low contact force, softness, fine manipulation, quiet operation, cleaning resistance, and long-term safety. Tendon-driven hands have strong theoretical advantages here, but only if they are robust and serviceable.

Which approach is better for factories?

Factories care about uptime, repeatability, service access, and known costs. In-finger or modular in-hand actuator systems may have an advantage if finger modules can be replaced quickly and thermal behavior is predictable. Tendon systems can still win if they survive duty cycles and reduce distal damage.

Can a robot hand combine both approaches?

Yes. Hybrid hands are likely. A hand can use remote tendons for some motions, local motors for others, passive compliance in selected joints, tactile fingertip modules, and replaceable skins. The market is already moving beyond simple labels.