BLOGS

For years, the actuator inside a humanoid or collaborative robot was treated as a trade secret. In 2025 and 2026 that changed: hardware teams started publishing CAD, bills of materials, test numbers and even friction models on GitHub. We read through the most active of these projects — the ones with real measurements rather than renders — and pulled out five lessons that matter if you are specifying joint modules for a robot program.

1. Torque density is decided at the air-gap radius, not by the gearbox

The clearest engineering post-mortem we found comes from an open-source quasi-direct-drive (QDD) actuator project last updated in June 2026. Its author documents three earlier versions that used small off-the-shelf drone BLDC motors — and admits they never produced enough torque density for dynamic, agile motion. The air-gap radius of a small motor is simply too short to generate the torque.

The fix was not a bigger gearbox. It was a larger-diameter "pancake" BLDC that maximizes the air-gap radius, combined with an internal cycloidal reducer to keep the package compact. The current prototype posts 10:1 reduction, 8.8 Nm peak torque, 22 rad/s maximum velocity and 12.6 rad/s nominal velocity at 24 V.

What this means on a spec sheet: when two modules advertise the same peak torque, ask about motor diameter, continuous torque at your duty cycle, and how heat leaves the stator. A small motor pushed to its limit and a larger motor running comfortably can post identical peak numbers and behave nothing alike after an hour of operation.

2. A lower reduction ratio buys behaviors a high ratio gives up

High-ratio gearboxes of 1:100 and above are excellent at multiplying torque and poor at everything else that matters for dynamic motion: friction and reflected inertia climb, backdrivability drops, and torque transparency disappears. For uneven terrain locomotion, hand-guiding, or any application where a person may push the robot, that mechanical impedance is a safety and control problem, not a rounding error.

The open-source QDD school — explicitly inspired by the MIT Mini Cheetah lineage — argues for a low ratio (10:1 in the project above) so the joint stays backdrivable and torque can be estimated from motor current through field-oriented control, with no additional torque sensor. That is proprioceptive sensing: the joint feels load through the same electronics that drive it.

This is not an argument against harmonic drives, which remain the right answer for high-torque, high-precision joints where holding torque and compactness dominate. The practical conclusion is architectural: most humanoid programs now run hybrid designs, using high-ratio harmonic gearing where the joint carries the body, and lower-ratio planetary or cycloidal stages where speed, transparency and responsiveness matter more. We compared these two architectures in detail in our harmonic drive vs planetary reducer guide.

3. Cycloidal reducers are moving out of the industrial catalog

Two developments are worth watching. First, an open-source ironless-rotor actuator published in late 2025 pairs a Halbach-array ironless rotor with a 3D-printed cycloidal-planetary gearbox and claims 29.4 Nm holding torque at roughly $40 for the actuator, or $70 including the controller — with zero-backlash intent. Second, commercial documentation for compact cycloidal joint modules and 8–42 mm planetary gearboxes is now published openly on GitHub, including application notes for humanoid, quadruped and robotic arm joints.

The build notes from the DIY project are the most valuable part. The author warns explicitly against PLA for the rotor, because the magnetic forces will creep the plastic over time, and recommends engineering filaments such as PA6-GF, PA6-CF or PET-CF, printed at 100 percent infill for the rotor. The gearbox runs on lithium grease, and the planet carrier needs tapped threads.

That is a free education in why "we will just 3D print it" stalls at the third revision. The kinematics work immediately; stiffness, creep and wear decide whether the joint is still accurate after a few thousand cycles.

4. Friction and backlash models are becoming a deliverable

The friction model in most simulators — Coulomb plus viscous — is too simplistic to represent real actuator behavior. It misses Stribeck effects, load dependence and quadratic friction terms, and the error compounds when you train a control policy in simulation and then deploy it on hardware.

An open-source project updated in September 2026 tackles this directly. It provides an identification pipeline to fit extended friction models from recorded trajectories, a library of already-identified models for common servos — including eRob80:50 and eRob80:100 harmonic-drive actuators — and an API to use them in MuJoCo CPU and MuJoCo Warp. It builds on an ICRA 2025 paper on extended friction models for physics simulation, and it sits alongside sim-to-real work from ETH Zurich that treats actuator identification as a first-class step, not an afterthought.

Why procurement should care: a robot's software is only as good as the joint data underneath it. When you shortlist a joint module supplier, ask for backlash and friction measured at your duty cycle, torque-speed curves at your actual bus voltage, thermal derating curves, and motion model files your simulation team can load. If those do not exist, your integration cost just went up.

5. The prototype-to-production gap is still where actuators die

Open-source projects prove concepts at $40 and a 3D printer. Production is a different discipline. The hard parts are not the kinematics but repeatability: consistent backlash across a batch, gear tooth and bearing life over millions of cycles, thermal derating inside a sealed limb with no airflow, encoder and cable routing through a moving joint, functional safety certification, and supply continuity for years after the design freeze.

This is the layer where lightweight joint module product lines differ from hobby builds. A production module combines a machined housing and precision transmission, validated grease and sealing, calibrated encoders, and a driver stack that has been characterized and life-tested — not a one-off assembly that works on the bench. It is also why the 2026 humanoid actuator trends that matter most to integrators are integration, repeatability and cost at volume, not laboratory peak numbers.

The shortlist: what to demand before you commit

Across all five lessons, the same questions keep deciding projects. Ask every supplier for:

• Continuous and peak torque, and how long the peak may be held
• Torque and speed quoted at the same operating point, at your bus voltage and current limit
• Backlash and friction data measured at your duty cycle, not a best-case table
• Radial, axial and overturning load ratings for the joint interface
• Thermal derating curves and expected service interval in the real enclosure
• Encoder resolution, communication protocol, and available CAD / URDF / MJCF models
• Service-life test evidence and batch consistency data from a representative sample

Our 10-point OEM buyer checklist for robot joint modules expands each of these into the exact questions to put in writing before an order.

Open-source projects referenced

• Quasi-direct-drive actuator, 10:1 cycloidal reducer, 8.8 Nm / 22 rad/s — github.com/JeongSeoJin/quasi-direct-drive-actuator (updated June 2026)
• Ironless-rotor QDD actuator, 3D-printed cycloidal-planetary gearbox — github.com/CKraft11/Ironless-QDD-Actuator (published late 2025)
• Extended friction models for servo actuators, with identified model library — github.com/Rhoban/bam (updated September 2026)
• Commercial documentation for cycloidal joint modules and 8–42 mm planetary gearboxes — github.com/SigGearDrive/SigGear-product-docs (updated September 2026)

The open-source wave is useful precisely because it publishes the failures as well as the wins. The teams writing it down are saving everyone else the third revision — and the buyers who read it arrive at the supplier conversation with better questions.

If you are sizing joints for a humanoid, collaborative arm or quadruped program and want a starting point, talk to Ti5ROBOT engineers with your duty cycle and envelope, and we will tell you which architecture fits — and what data we have to back it.