In this guide
How Tendon-Driven Robot Arms Work: Cables, Compliance and Learning from Demonstration
A tendon-driven robot arm moves its joints with cables instead of motors. The motors sit in the robot's torso and pull thin braided lines — tendons — that run through guide tubes to each joint, the way the muscles in your forearm pull tendons that move your fingers. The moving arm itself carries almost no motor weight, so it is light, it yields instead of pushing through when it touches something, and it can work near people without the heavy, rigid arm that a conventional servo arm needs. This guide explains the mechanics, the safety argument, what "learning from demonstration" actually means at student scale, and how to build one on a bench.
The news that put tendon arms on the map
On 6 November 2026 the Swiss startup Embodied AI unveiled AIR at Humanoid Europe 2026 in Munich: a wheeled humanoid — torso, two arms and a head on a wheeled base, deliberately no legs — whose arms are tendon-driven. The company reported 98% reliability on a cable-handling evaluation against a 90% target, and said the robot learns factory tasks from human demonstration, with a remote operator running the first shifts while every run is recorded as training data. It also said the design follows ISO/TS 15066:2016, the collaborative-robot safety specification, with CE marking planned for 2027. All of those figures are company-reported, not independently verified.
AIR was not the only wheeled humanoid in the news that week. On 9 November 2026 Boston Dynamics released footage of its Atlas humanoid fitted with wheels instead of feet, landing a 360-degree jumping spin in a research test — a research platform, not a product. And on 4 November 2026 Apple's research team published HALO, a system that trains humanoid motion in simulation and runs it on a Unitree G1 robot in real rooms, reporting 91% average success across 360 single-task tests. Three different stories, one pattern: 2026 is the year humanoids stopped trying to walk everywhere and started rolling, and the interesting engineering moved into arms, safety and learning.
Students should read AIR the way the trend-watch project does: not as a machine to copy, but as an idea to translate. A desktop tendon-driven arm that learns a cable-handling task by demonstration is the honest student-scale version, and the full build is published as a project on this site: the AIR-inspired tendon-driven wheeled humanoid robot (teach-and-replay arms, Wi-Fi dashboard, cable tray with three guide clips). The rest of this guide is the physics and method behind that build.
What a tendon drive actually is
A tendon is a cable that works in tension only: it can pull, never push. One end attaches to a servo horn in the torso, the line runs through a guide tube to the joint, and the other end attaches to a lever arm on the joint. When the servo winds the line in, the joint rotates one way; a second line (the antagonist) or a spring returns it the other way. The torque at the joint is simply tension multiplied by the lever radius, so a small servo pulling hard on a short lever gives the same torque as a bigger servo on a longer one.
That is the whole trick, and it changes the arm in three ways. First, weight: a conventional arm puts a servo at every joint, so the shoulder servo lifts the elbow servo, which lifts the wrist servo, which lifts the gripper. A tendon arm puts all the servos in the torso, so the moving arm carries only aluminium links, cables and a gripper. Second, inertia: a light arm accelerates and stops faster, and when it hits something it carries far less energy — impact energy scales with mass times speed squared, so halving the moving mass and capping the speed cuts the impact hard. Third, shape: the arm can be slim, because the bulky parts live in the torso where there is room for them.
The cost is complexity you can feel on the bench. A cable that only pulls needs a return path for every direction of motion. The line slides inside its guide tube, so friction eats some of the servo's torque and the joint angle is no longer exactly the servo angle. Braided line stretches a little under load and creeps over weeks, so taught poses drift and the tendons need re-tensioning. And the tube itself must be fixed at both ends with gentle bends, because a sharp bend multiplies friction and a loose tube lets the line saw into printed joints.
| Tendon-driven arm | Rigid servo arm | |
|---|---|---|
| Motors | In the torso; cables drive the joints | One servo at every joint |
| Moving mass | Links, cables and gripper only | Full motor weight on the arm |
| On contact | Tendons go slack; arm yields | Rigid arm holds position |
| Torque | Tension × lever radius; traded for lightness | Direct from the servo at the joint |
| Failure modes | Stretch, slack, friction, wear | Stripped gears, stalled servo |
| Maintenance | Re-tensioning and re-teaching | Gear and servo replacement |
Why tendons make arms safer near people
Conventional robot safety is mostly about keeping people away: fences, light curtains, interlocked gates. Collaborative robots invert that: the robot is allowed near people, so the robot itself must be incapable of hurting them. ISO/TS 15066:2016, the technical specification AIR's design follows, sets out how — power and force limiting, speed and separation monitoring, and pressure limits for different body regions. It is a specification, not a certificate you can buy for a bench build, and no student prototype can claim compliance with it.
Tendons help the power-and-force-limiting case in a way that is easy to miss. The force an arm can apply in a collision comes from its moving mass and its speed. A rigid arm carries its motors, so even a slow collision carries the full weight of the arm. A tendon arm carries almost nothing, so the same collision carries far less energy, and the tendons add a second safety property for free: when the arm meets resistance, the cable goes slack instead of transmitting the servo's full stall torque into the obstacle. A servo current limit and a joint-speed cap — about 40 degrees per second in the student build — add two more layers, and an ultrasonic presence zone that pauses motion plus a latching e-stop complete the demonstration safety behaviour. None of this is certified safety. It is the honest student version of the same idea: make the arm light, make it yield, make it slow, and make it stoppable.
Learning from demonstration: what the words actually mean
"Learns from demonstration" is doing a lot of work in the AIR coverage, so it is worth separating three things that share the phrase. The first is teach-and-replay: a human moves the arm through a task while the controller records the joint angles, and the robot later plays those angles back. That is joint-angle playback. It learns nothing, generalises to nothing, and needs a new recording for a new task. It is also the only version a student can build in a semester with an ESP32, and it is genuinely useful — it is how the student build teaches its cable-handling task.
The second is imitation learning, sometimes called behaviour cloning: the recorded demonstrations become training data for a model that maps what the robot sees to what its joints should do, so the robot can handle variations it was never shown exactly. That needs a dataset of demonstrations, a vision or state input, and training compute — a real machine-learning project, and a big one. The third is reinforcement learning, where the robot improves by trial and error against a reward signal, usually first in simulation and then transferred to the real robot — the sim-to-real approach in Apple's HALO paper, which trained motion in simulation and ran it on a real Unitree G1.
AIR's own description sits between the first and the second: a remote operator runs the first shifts, every run is recorded as training data, and the company says the robot takes over autonomously within about a month. The honest student ladder is: build teach-and-replay first and log every attempt honestly, then — only if the course has room for it — record a dataset of demonstrations and train a small behaviour-cloning model on it. Calling playback "learning" is the single most common honesty failure in student robotics write-ups, and reviewers notice it immediately.
Building one on a bench: the parts and the pitfalls
A bench tendon arm needs few parts, and every one of them is doing a specific job. The controller is an ESP32 with a PCA9685 16-channel PWM driver, because six servos need six clean PWM signals and the ESP32's own pins are better spent on the dashboard and sensors. The shoulder and elbow drives are MG996R servos (stall torque 9.4 kg·cm at 4.8 V, 11 kg·cm at 6 V, stall current 2.5 A at 6 V — datasheet values), and the wrist and gripper use SG90 micro servos (1.8 kg·cm at 4.8 V, 9 g — datasheet values). The tendons are braided Dyneema fishing line through PTFE guide tubes; braided line beats monofilament because it does not coil and it handles the repeated bending better. The presence sensor is an HC-SR04 ultrasonic module (2–400 cm range, datasheet value) watching a zone of about 30 cm, and the head carries an ESP32-CAM so the work area is visible on the dashboard.
Four pitfalls decide whether the build works or frustrates its builder. First, power: the MG996R can pull 2.5 A at stall, so the servos need their own 5–6 V rail with real current capacity — never power them from the ESP32's 5 V pin, which will brown out and reset the board mid-task. Second, routing: fix each PTFE tube at both ends, keep bends gentle, and tension the line so the joint holds position without slack but is not guitar-string tight. Third, calibration: drive each joint to its limits, match the dashboard slider range to the safe joint range, and set the speed cap in firmware before the first replay. Fourth, expectations: the gripper handles about 100–200 g (design target), tendon tension falls with reach, and the SG90 wrist is the weakest link — heavier cables stall the gripper or miss the socket. The full parts list, wiring, calibration procedure and honest limits are in the AIR-inspired tendon-driven wheeled humanoid project, which is the build this guide explains.
The record loop itself is short, and it is the whole of teach-and-replay. In teach mode the firmware samples the six joint angles about ten times a second and appends them to a named task file; in replay mode it reads the file back and drives the servos through the recorded angles at the capped speed, while the dashboard ticks off each step of the cable task. A minimal version of that loop looks like this:
// Teach mode: sample joint angles at ~10 Hz into the task file
if (teaching) {
for (int j = 0; j < 6; j++) {
taskFile.print(jointAngle[j]);
taskFile.print(j < 5 ? "," : "\n");
}
delay(100); // ~10 Hz sample rate
}
That is all the "learning" the student build does, and the listing says so plainly: the robot plays back recorded joint angles and learns nothing from the task. Students who want to go further should read the inverse kinematics guide for robot arms before adding any position control, and the ROS 2 basics guide before considering a framework larger than the ESP32 dashboard.
Frequently asked questions
Do tendon arms lift less than rigid arms?
Yes, at the same servo size, because the tendon trades torque for lightness: the torque at the joint is tension times lever radius, and the line, the tube friction and the small lever all cost torque. The student build lifts about 100–200 g at the gripper (design target), far below AIR's reported 5 kg arm lift. The trade is deliberate — lightness and yielding are the point, not maximum payload.
Can I use ordinary fishing line for the tendons?
Braided Dyneema fishing line works well and is what the student build uses; monofilament is a poor choice because it coils, stretches more and kinks at the guide-tube entries. Whatever line you use, expect to re-tension after the first hours of use as the braid beds in, and re-teach any poses that drift.
Is teach-and-replay machine learning?
No. It records joint angles and plays them back, with no model, no training data and no generalisation. A new cable route needs a new taught recording. Imitation learning starts where teach-and-replay stops: the demonstrations become training data for a model. The guide keeps the two separate because reviewers do.
Do I need force sensors to make the arm safe?
Not for a bench demonstration, where light mass, tendon slack, a speed cap, a presence pause and an e-stop are the safety behaviour. But those are demonstration behaviours, not certified safety: there is no force sensing, no safety-rated controller and no risk assessment, so the robot must never work beside a person unsupervised. Force sensing is the first upgrade if the project ever moves beyond a demo.
Why a wheeled base instead of legs?
AIR's answer and the student's answer are the same: wheels are simpler, more stable and cheaper than legs, and the task — handling cables on a bench — does not need walking. A two-wheel driven base with a castor is a bench trolley, not a factory mobile robot, and the guide says so wherever the base is described.
Limitations
- This guide explains principles and a bench build, not a certified collaborative robot. Nothing here is ISO/TS 15066 compliance, and the student build must never be presented as certified human-safe.
- All AIR figures in this guide — the 6 November 2026 unveiling, the reported 98% reliability, the one-month autonomous takeover, the 2 kg arm lifting 5 kg — are company-reported or investor claims, not independently verified, and the guide labels them as such.
- Tendon arms need maintenance: braided line stretches, PTFE tubes wear, taught poses drift as tension changes, and tendons need re-tensioning and re-teaching after heavy use.
- The student build handles one fixed bench task — a cable, three clips and a socket on a tray. It cannot fetch a cable from outside the tray, and its wheeled base is a bench trolley.
- The physics in this guide is first-principles reasoning for a student build. Servo torque, tendon tension and impact energy all depend on the actual parts and geometry used, so treat the figures as design targets and measure your own build.
The tendon arm is worth building because it makes three ideas testable in one project: that moving mass out of the arm changes both performance and safety, that a cable can be a precision transmission if you respect friction and stretch, and that "learning from demonstration" has an honest ladder from playback to imitation. The AIR-inspired tendon-driven wheeled humanoid project is the bench version of all three — and the place to start if you want to build one.