Robot Dog

Project Goal

Design, build, and program a highly dynamic 12-motor quadruped (~8 kg) to gain hands-on experience with complex system design, mechanical fabrication, and electronics selection.

Actuator Selection (September 2026)

An actuator is arguably the most important part of a quadruped robot and is also one of the biggest drivers of final cost.

QDD Actuators

Typically, robots of this size use a low-KV brushless motor driven by a Field Oriented Controller (FOC) paired with a high reduction gearbox (around 8:1). This type of actuator is also known as a quasi-direct-drive (QDD) actuator. QDD actuators are back drivable and have high torque transparency which allows the robot to absorb impacts exceptionally well and sense force through the motor without the need for additional sensors. While high-torque servos can be good for very small quadruped robots, they make for a significantly less capable robot at this scale.

After examining different options for QDD actuators I decided to pursue a capstan based system for the following reasons:

Off-the-shelf QDD actuators are relatively affordable and very capable, which is why they are used widely on industrial robots, however, using premade actuators eliminates a major part of the design challenge of this project.
Designing and building a standard QDD actuator would be prohibitively expensive for this project due to the high costs associated with CNCing precision cycloidal or planetary gearboxes and housings. 3D-printed gearboxes are not viable because of high friction, low strength, high backlash, and short lifespan.
Capstan drives can be 3D printed and don't suffer from most of the problems associated with a 3D-printed gearbox because they use rope to transmit power instead of gear teeth.

Capstan Drives

A capstan drive operates by wrapping a rope multiple times around a small motor-driven drum, which then winds and unwinds onto a larger drum to drive rotation. The friction from the multiple wraps prevents the rope from slipping, even under high loads, while effectively distributing the applied force over a large surface area.

While capstan drives have been successfully implemented in other quadruped projects — such as Ahead.io’s Stanley and Aaed Musa’s Cara — a major drawback evident in both systems is their overall bulk.

To prevent excessive rope wear, the small drum's diameter must be significantly larger than the rope's diameter. This dictates a proportionally massive large drum to achieve the desired reduction ratio. This inherent bulk is further exacerbated by the fact that the larger drum must be significantly offset from the central motor shaft.

Capstan Drive Key Parts

Linear Capstan Drives

A linear capstan drive operates on the same fundamental principle as a standard capstan drive: friction between the rope and the small motor drum prevents slipping while distributing the load across a large surface area. However, rather than wrapping around a second, larger drum, the rope runs in a straight path parallel to a linear rail. This configuration effectively converts rotational motion directly into linear motion.

Because this actuator necessitates a drastically different leg architecture compared to a standard quadruped robot, I prototyped a complete leg utilizing linear capstan drives to better evaluate the system's potential advantages and limitations.

Linear Capstan Drive

Linear Capstan Based Leg Geometry

This leg design represents the simplest method I have found for integrating linear capstans into a quadruped robot. By utilizing two linear actuators joined at the foot, the system can achieve any foot position within a 2D plane simply by varying their respective lengths. When paired with an abduction/adduction (ABAD) joint, the leg will be capable of reaching every position required for full 3D walking.

This architecture eliminates the need for a complex coaxial hip pivot and yields a profile significantly thinner than a traditional capstan-driven leg. Structurally, the design is highly efficient; because most forces travel axially along the actuators, the components can remain exceptionally thin and lightweight, drastically reducing the leg's overall moment of inertia. Furthermore, the absence of a gearbox results in outstanding force transparency, while the dual-actuator setup distributes the load across both motors evenly.

Linear Actuator Configuration

Prototype Leg

Linear Capstan Drive Test Stand (September 2026)

The primary goal of this isolated test rig was to empirically validate the linear capstan geometry and verify that the synthetic rope mechanism could survive high-tension, high-cycle operation before integrating it into a complete leg assembly.

Motor Selection & Field Oriented Control

Motor: Selected an Eaglepower LA8308 rated at 90 kV (measured at ~100 kV) for high torque-to-current ratio and good balance of performance and cost.
Controller: Moteus r4.11 controller which is mounted to the back of the stator. The Moteus ecosystem is explicitly designed for high-bandwidth torque control, making it ideal for this application.
Absolute Position Tracking: Designed a custom 3D-printed adapter to mount a diametric magnet to the motor shaft. This interfaces with the controller's onboard magnetic encoder, providing the exact rotor positioning necessary to utilize FOC.

Soldering Phase Wires

Full Assembly

String Termination & Endstop

System Specifics and Construction

Drum Sizing: The 28mm drum diameter was mathematically sized based on the estimated 8kg vehicle weight and the peak loads required for aggressive maneuvers (e.g., jumping with only two legs contacting the ground). This geometry yields a kinematic ratio where three full motor revolutions produce a 264mm stroke. The 3D printed drum is reinforced with 55mm M3 screws to prevent shearing along layer lines.
String Routing: The drive utilizes 2mm Dyneema synthetic rope, which was pre-stretched prior to installation to minimize mechanical settling. It is wrapped around the drum six times and permanently anchored at the center to prevent slipping during assembly. Under active load, friction dynamics governed by the capstan equation easily prevent the rope from slipping.
Rope Termination & Tensioning: To avoid placing extreme point loads on the plastic slider assembly, the Dyneema string terminates by wrapping around two M4 bolts to distribute the force. It is then routed into an integrated tensioner adjustable via two bolts.
Physical Homing: Hard stops are integrated at each end of the sliding assembly. These stops are designed to conform directly to the string wrapped around the main drum, providing a rigid, repeatable physical limit used for motor homing upon power-up.
Construction: 3030 aluminum extrusions were used to make a rigid frame and robust linear rails and bearings were used to enable long-term testing of the rope. 3D-printed brackets mount the motor, controller, and drum.

Endurance Testing

Testing & Results

Cycle Testing: To date, the testing rig has successfully completed over 100,000 continuous cycles with no noticable abrasion of the rope.
Rope Elongation: The rope elongated slightly but did not become criticaly loose so the actuator lost some precision. When the actuator was put under a high load, the rope would press into the threads on the M4 bolts near the termination, causing further loosening of the rope.
Force: The actuator achieved a maximum force of approximately 50 N which is significantly less than needed for the final robot but about what was expected given the 24V 5A power supply.

Complete Actuator — Test Leg V1 (October 2026)

After successfully validating the rope kinematics on the test stand, I designed a complete leg to test an actuator better suited for the final robot. Following several iterations, I finalized the current design.

Similar to the capstan test stand, this system utilizes 28 mm drums, 90 kV Eaglepower motors, and r4.11 controllers.

The actuators are positioned as closely together as possible to prevent them from splaying apart when fully retracted. Lightweight carbon fiber tubes replace traditional linear rails, attaching to 3D-printed blocks at the top and bottom that double as string termination points.

Before committing to building the entire leg, I prototyped a single actuator—essentially the complete leg split in half.

Complete CAD — Leg V1

Rigging and Drive Assembly

Rigging & Anchoring: Upgraded to inextensible 2.2 mm Dyneema DM20 rope (660 N WLL) to eliminate elongation, routing it through large-radius bends to a screw tensioner to prevent thread-induced pinching.
Drive Assembly: Mounted motors tightly on a heat-dissipating aluminum plate for maximum compactness, and secured controllers to heat sinks to maintain precise encoder spacing and support continuous 20 A loads.

Rope Anchor

Motor and Controller Mounts

Linear Bearing Hub

Bearing Selection: Utilizes 30 mm steel ball bearings for coaxial rotation with the motor drum, and LM8UU polymer bearings for linear movement. Polymer is required here, as steel linear bearings are incompatible with the carbon fiber actuator tubes.
Hub Design: Features a split, two-part hub that securely sandwiches the LM8UU linear bearings inside. This split design enables support-free 3D printing, which is critical for achieving a precise fit with the bearings.

Linear Bearing Hub with Bearings

Testing and Results

Nominal Performance & Rigging: Guided by a vertical linear rail, the actuator operated smoothly and quietly under purely vertical loads. The inextensible rope and updated anchoring design succeeded; under high load, the slack side safely lost tension while the load-bearing side maintained its shape without noticeable elongation.
Abduction Load Limitations: When angled to simulate a quadruped's abduction joint, the relatively short polymer bearings struggled against the high torque exerted by the long carbon fiber tubes. This induced severe, friction-based binding that nearly seized the actuator's linear movement entirely.
Endurance & Test Termination: Testing concluded at 10,000 cycles due to power supply wire fatigue and the high friction encountered at abduction angles, though the polymer bearings exhibited no visible wear.
Next Steps: To resolve the friction binding, the next actuator iteration will transition to a single 2020 aluminum extrusion guided by V-groove ball bearings. This new architecture will eliminate the binding issue while also making the overall design significantly more assembly-friendly.

Final Actuator

Actuator Testing

← Projects

Tension Link Load Cell →

Page updated

Google Sites

Report abuse