AIR-Inspired Tendon-Driven Wheeled Humanoid Robot that Learns Cable-Handling Tasks from Human Demonstration

This project builds a tendon-driven wheeled humanoid robot that copies the idea behind AIR, the wheeled humanoid unveiled by Swiss spin-out Embodied AI at Humanoid Europe 2026 in Munich on 6 November 2026: two arms that handle cables in electronics assembly, taught by watching a human operator. Cable handling — picking a cable, routing it through clips and seating its connector — is repetitive bench work that is hard to automate with rigid arms, because a heavy arm moving fast near people is a safety risk and a heavy arm needs a heavy base. AIR answers that with tendon-driven arms: the motors, gearboxes and electronics sit in the torso, thin cables drive the joints, and the moving arms are almost massless, so they yield on contact instead of pushing through it. The company reported 98% reliability on a cable-handling evaluation against a 90% target. This student build translates that idea honestly to a desktop robot: two 3-DOF tendon-driven arms on a wheeled torso, taught by hand-guiding and replayed from recorded joint angles, with a presence pause and a latching e-stop.

The end-to-end mechanism runs teach, record, replay and safety in one loop. In teach mode the student moves each arm by hand or from the dashboard sliders, and the ESP32 samples all six joint angles at about 10 Hz into a named task file, storing the cable-handling sequence as a list of joint positions. In replay mode the ESP32 reads the file back and drives the six servos through the PCA9685 driver at the capped joint speed, so the arm repeats the taught path: pick the cable from the tray, route it through the three guide clips in order, and seat its connector in the socket. While it moves, the HC-SR04 watches the presence zone and the firmware watches servo current: a hand within about 30 cm or a sudden load pauses the motion and lets the tendons go slack. The ESP32-CAM head streams the work area to the dashboard, where the checklist, replay progress and safety status update live.

1. The builder assembles the robot: a two-wheel driven base with a castor, a torso column, two 3-DOF arms of aluminium links, and the MG996R shoulder and elbow servos plus SG90 wrist and gripper servos mounted in the torso, with the ESP32, PCA9685 driver and HC-SR04 wired on the base.
2. The builder routes the tendons: braided Dyneema line runs from each torso servo through a PTFE guide tube to its arm joint, is tensioned so the joint holds position without slack, and is tied off so a pull on the line moves the joint and a release lets it go slack.
3. The builder calibrates the motion: each joint is driven to its limits, the dashboard slider range is matched to the safe joint range, the joint-speed cap near 40 degrees per second is set in firmware, and the HC-SR04 presence zone is marked at about 30 cm.
4. The builder teaches the cable task: in teach mode the arm is hand-guided or jogged from the dashboard to pick the cable from the tray, route it through the three guide clips in order, and seat its connector in the socket, while the ESP32 records the joint angles at about 10 Hz into a named task file.
5. The builder replays the task: replay mode drives the servos through the recorded angles at the capped speed, the dashboard checklist ticks off each clip and the socket step, and the task-success log records whether the connector seated.
6. The builder demonstrates the safety behaviour: a hand is moved into the presence zone mid-replay to show the motion pausing, the e-stop is pressed to show motion staying halted until manual reset, and a light push on the arm shows the tendons going slack instead of resisting.
7. The builder exports the task as CSV, points to the recorded angles, the clip steps, the pause events and the e-stop event, and reports honestly in the report which replays seated the connector and which did not.

Tech stack:

• ESP32 · main controller and Wi-Fi dashboard host
• PCA9685 servo driver · 16-channel 12-bit PWM over I2C
• MG996R servos · shoulder and elbow tendon drives
• SG90 micro servos · wrist and gripper drives
• Dyneema tendons · braided line through PTFE guide tubes
• ESP32-CAM · head camera with OV2640 sensor
• HC-SR04 ultrasonic sensor · presence zone sensing
• Wi-Fi dashboard · joint control and task log and CSV export

This tendon-driven wheeled humanoid earns its place on a review table because it makes three ideas testable in one build. First, the tendon drive: MG996R servos mounted in the torso pull braided Dyneema tendons through PTFE guide tubes, so the moving arms carry almost no motor weight and go slack instead of staying rigid when they meet resistance. Second, demonstration learning in its honest student form: in teach mode the student hand-guides each joint or jogs it from the dashboard while the ESP32 records joint angles at about 10 Hz, and in replay mode the robot repeats the recorded cable-handling sequence on its own. Third, human-safe behaviour built into the motion: a joint-speed cap near 40 degrees per second, a servo current limit, an ultrasonic presence zone that pauses motion, and a latching e-stop. The Wi-Fi dashboard ties it together with live joint angles, the recorded-task list, replay progress and a CSV export of the task log.

• [Tendon-driven arms with torso-mounted servos] The MG996R shoulder and elbow servos sit in the torso and pull braided Dyneema tendons through PTFE guide tubes to the arm joints, so the moving arms carry almost no motor weight and the arm trades torque for lightness — implemented in the base build.
• [Teach-and-replay demonstration learning] In teach mode the student hand-guides each arm joint or jogs it from the dashboard sliders while the ESP32 records joint angles at about 10 Hz into a named task file, and in replay mode the robot repeats the recorded sequence on its own — implemented in the base build, and stated plainly as joint-angle playback, not machine learning.
• [Cable-handling demo task] The taught task is the same task class AIR was evaluated on: pick a cable from the tray, route it through three guide clips in order, and seat its connector in the socket, with the dashboard checklist tracking each step — implemented in the base build.
• [Human-safe yielding on contact] The tendons go slack when the arm meets resistance, the servo current limit stops a stalled drive, joint speed is capped near 40 degrees per second, and the HC-SR04 presence zone pauses motion when a hand comes within about 30 cm — implemented in the base build as demonstration safety behaviour.
• [Latching e-stop and capped joint speed] A red mushroom e-stop button cuts all drive power until it is twisted back to reset, and the firmware holds every joint under 40 degrees per second, so the arm never moves at full servo speed near a person — implemented in the base build.
• [Wi-Fi dashboard with task log and CSV export] The dashboard shows live joint angles for both arms, the recorded-task list, replay progress, safety status and a task-success log, and exports the recorded task as CSV for the report — implemented in the base build.
• [ESP32-CAM head camera] The head carries an ESP32-CAM so the work area can be watched from the dashboard while a task is taught or replayed, using the OV2640 sensor at 2 megapixels (Datasheet value) — implemented in the base build.
• [Force sensing and compliant torque control] Future scope only, not implemented: torque or force sensors at the joints that would let the arm measure contact force instead of inferring it from servo current, listed as an extension rather than a claimed feature, since the base build has no force sensing.

• Complete fabricated wheeled torso robot (two-wheel driven base with castor, torso column, two 3-DOF tendon-driven arms, 2-finger grippers, cable tray with three guide clips and a socket)
• Electronics set (ESP32 board, PCA9685 servo driver, MG996R and SG90 servos, ESP32-CAM head camera, HC-SR04 presence sensor, latching e-stop button, 12 V adapter with 5 V and 6 V servo rail)
• Teach-and-replay firmware and Wi-Fi dashboard (joint sliders, teach and record, replay, task checklist, safety status, task-success log, CSV export)
• Setup and calibration guide (tendon routing and tensioning, servo wiring, joint-limit calibration, presence-zone marking, e-stop test, CSV reading worksheet)
• Project report PDF (tendon-drive background, AIR comparison limits, teach-and-replay design, safety design, demonstration results, honest measurement limits)
• PPT presentation for final review
• Viva Q&A preparation document (tendon vs rigid arms, teach-and-replay vs machine learning, safety design, AIR comparison limits, payload limits, single-bench-task limits)