Quasi-direct drive

Quasi-direct drive (commonly abbreviated QDD) is an actuator architecture that pairs a high-torque-density electric motor, typically a large-diameter brushless DC motor, with a low-ratio mechanical transmission, most often a single-stage planetary gear reduction. Because the gear ratio stays low, commonly cited as single digits up to roughly 10:1 and only occasionally as high as about 20:1, a QDD joint remains highly backdrivable and its output torque can be estimated directly from motor current, without a dedicated torque sensor. ^[1] This property, generally called torque transparency, has made QDD the dominant actuator choice for research quadrupeds and a growing share of commercial legged robots, and it sits at the center of an ongoing design debate in humanoid robot engineering against high-reduction transmissions such as the harmonic drive and cycloidal drive. ^[2]

In brief

Most electric motors spin fast but twist weakly, so robot joints usually add a gearbox to trade speed for torque, the same way a low gear on a bicycle makes a hill climbable at the cost of pedaling speed. A conventional robot gearbox multiplies torque by 50 times or more, which is excellent for lifting heavy loads, but that much mechanical amplification also buries the feedback the motor would otherwise feel from the outside world, similar to judging how hard you are squeezing something while wearing thick oven mitts. Quasi-direct drive uses a much smaller gearbox, typically in the single digits up to about 10:1, so the motor keeps enough of a direct mechanical link to the joint that its electrical current still reveals, with reasonable accuracy, how hard that joint is pushing or being pushed on. Engineers can then skip a separate force sensor at the joint, and the joint gets a natural give when it meets an unexpected obstacle, a person, or the ground, a trait that matters for walking robots and for machines that learn by trial and error.

How it works

Torque transparency: sensing force through current

Brushless motor controllers typically use field-oriented control, which measures the motor's phase currents and resolves them into a torque-producing (quadrature) component and a flux-producing component. For a well-characterized motor, the torque-producing current is proportional to output torque through the motor's torque constant, so a controller can estimate shaft torque from current alone. ^[1] In a low-ratio, low-friction transmission that estimate stays reasonably accurate all the way to the joint, a property called torque transparency. In a high-reduction transmission, by contrast, internal friction and elastic windup absorb and distort a large share of the torque before it reaches the output, so current sensing alone becomes unreliable and designers commonly add a dedicated force/torque sensor at the joint instead. ^{[3] [4]} Researchers have also built learned "actuator network" models that correct for the small nonlinearities that remain even in QDD transmissions, further narrowing the gap between the current reading and true output torque. ^[5]

Backdrivability and reflected inertia

A joint is backdrivable if an external force applied at the output can move it with little resistance, ideally approaching zero added torque. ^[6] The obstacle to backdrivability is reflected inertia: everything on the load side of a gearbox appears at the motor shaft scaled by the square of the gear ratio, so a 50:1 reduction makes the load's inertia and friction feel 2,500 times larger to the motor than they actually are. ^{[1] [6]} A QDD transmission's low ratio keeps that multiplier small, so external pushes, ground impacts, or a person's touch reach the motor with comparatively little distortion, and the joint can be driven backward by hand or by contact forces almost as easily as an unpowered joint would move. ^[6] Backdrivability is also why QDD joints tend to absorb shock rather than transmit it into gear teeth, a property that matters for repeated foot strikes in running and jumping robots. ^{[1] [7]}

Motor selection: large diameter, modest speed

Because a QDD transmission multiplies torque only a little, most of the required torque has to come from the motor itself rather than from gearing. Motor torque scales roughly linearly with the length of the magnetic stack but with the square of the air-gap radius, so designers get more torque per added mass by making the motor wider rather than longer, producing the flat, wide "pancake" rotor shape typical of QDD actuators. ^{[8] [9]} These motors also tend to sit toward the high-torque, lower-speed end of the brushless motor spectrum: a motor's speed constant (KV, in RPM per volt) and torque constant (Kt) are inversely related, so a motor wound for a lower KV produces more torque per amp of current at the cost of top speed, a trade that suits a transmission that will not multiply torque much further downstream. ^[10] Many early QDD actuators, including MIT's, started from rewound hobbyist or drone motors before purpose-built versions became available. ^{[1] [9]}

Gear ratio and design variants

Sources disagree on the exact boundary of "quasi-direct," which is a descriptive term rather than a strict engineering standard. Overview sources place the range anywhere from about 2:1 to 15:1, other technical treatments cite 3:1 to 10:1, and some legged-robot literature extends the label up to about 20:1 to 25:1 for the most torque-hungry designs. ^{[1] [11] [12]} The practical ceiling is set partly by geometry: a single planetary stage can only pack so many teeth into a given envelope, and comparative actuator-design research finds internal single-stage planetary gearboxes work best in roughly the 5:1 to 7:1 range while external single-stage designs extend usefully to about 7:1 to 11:1 before a second stage becomes necessary. ^[13] The MIT Cheetah 2 and Cheetah 3 quadrupeds used single-stage planetary reductions of 5.8:1 and 7.67:1 respectively, and the open-source Mini Cheetah actuator uses about 6:1. ^{[14] [15]} At the upper edge of the range, some humanoid joints use a two-stage planetary reduction to reach roughly 20:1 while still being described as QDD; an independent teardown of Unitree's G1 humanoid calculated a total reduction near 20.6:1 from a two-stage planetary gear train (stages of approximately 3.3:1 and 6.25:1 based on measured tooth counts). ^[16] All of these figures sit far below the reduction ratios used in high-reduction transmissions, discussed next.

Scroll sideways for more →

Quasi-direct drive versus high-reduction transmissions

The alternative to QDD is a high-reduction transmission, chiefly the strain-wave (harmonic) drive and the cycloidal drive, which multiply torque far more aggressively at the cost of the transparency and backdrivability that QDD preserves. Harmonic drives typically offer reduction ratios from about 30:1 up into the low hundreds to one, with many catalog product lines spanning 30:1 to 160:1 and some specialized series reaching 320:1. ^{[17] [18]} Cycloidal reducers span a comparable or wider range: single-stage catalog units commonly run from about 6:1 to 119:1, and the "RV reducer" style used in heavy industrial joints often covers roughly 20:1 to 225:1. ^[19] Both use their high reduction to reach torque densities that QDD cannot match at a given size. By one set of engineering comparisons circulated among actuator suppliers, harmonic drives deliver roughly 25 to 39 newton-meters of torque per kilogram of actuator mass, versus roughly 7 to 28 Nm/kg for planetary QDD designs, though such figures vary with the specific product and measurement convention and should be read as illustrative rather than definitive. ^[20]

The trade is not one-directional. QDD's efficiency (planetary gear meshes commonly exceed 90 to 95 percent) and low internal friction reduce heat and battery drain, and its backdrivability lets a joint absorb impact rather than transmit a shock spike into gear teeth, which is valuable wherever a limb repeatedly strikes the ground or an unplanned object. ^{[7] [20]} Harmonic drives, by contrast, are prized for near-zero backlash and high positioning precision but are comparatively fragile under impact because their flexspline relies on repeated elastic flexing, and they are generally not backdrivable, often self-locking at higher ratios. ^{[2] [20]} Cycloidal drives sit in between: their rolling-lobe contact spreads load over a larger area than harmonic drives, giving better shock tolerance, but at typical robot ratios they still resist backdriving and usually need their own torque sensor for accurate force control. ^{[2] [19]}

Scroll sideways for more →

Neither approach is objectively superior; each fits different parts of a robot and different design priorities. A humanoid built for precise, controlled manipulation or heavy static loads leans toward high reduction, while one built for dynamic locomotion, safe contact, or reinforcement-learning-based control leans toward QDD. In practice, Tesla's Optimus illustrates the high-reduction bet: its rotary joints (shoulders, hips, knees, elbows) combine a frameless torque motor with a harmonic drive and an explicit torque sensor, and its linear joints use a planetary roller screw, prioritizing force density and positioning precision. ^{[22] [23]} Unitree's quadrupeds and its Unitree G1 humanoid illustrate the opposite bet, using QDD-style actuation to cut cost and preserve compliance and dynamic agility, accepting somewhat lower torque density and more gear wear in exchange. ^{[16] [24]} Some designs blend both: a two-stage planetary QDD joint like G1's knee sits deliberately near the boundary between the two camps.

Origins: from the MIT Cheetah to today

The QDD design philosophy traces most directly to the MIT Biomimetic Robotics Lab, led by mechanical engineering professor Sangbae Kim, whose Cheetah quadruped project began in 2009. ^[25] The lab's early work, published by Sangjae Seok, Albert Wang, Meng Yee Michael Chuah, David Otten, Jeffrey Lang, and Kim, laid out design principles for what they termed proprioceptive actuation: pairing a high-torque-density motor with a low, efficient gear reduction so the robot could sense and control ground-contact forces through the actuator itself, without dedicated force or torque sensors at the feet. ^[14] Patrick Wensing, Albert Wang, Sangok Seok, David Otten, Jeffrey Lang, and Sangbae Kim formalized this further in a 2017 IEEE Transactions on Robotics paper, "Proprioceptive Actuator Design in the MIT Cheetah," reporting an actuator able to deliver more than 26 newton-meters of continuous torque at speeds beyond 37 radians per second while remaining highly backdrivable for impact mitigation and high-bandwidth force control. ^[1] The Cheetah 2 and Cheetah 3 robots used single-stage planetary reductions of 5.8:1 and 7.67:1, with Cheetah 3 extending the approach to full three-dimensional per-leg force control without any dedicated joint force or torque sensors. ^{[15] [26]}

In 2019, the lab open-sourced a smaller, cheaper version of the same actuator concept in the Mini Cheetah, a 12-degrees-of-freedom quadruped with modular joint actuators (roughly 6:1 reduction, about 17 Nm peak and 6.9 Nm continuous torque) that became a widely replicated research platform and the first quadruped robot documented performing an untethered backflip. ^[27] The lab later carried a higher-power version of the same actuator philosophy into a bipedal research platform, sometimes referred to as the MIT Humanoid, aimed at dynamic and acrobatic whole-body motion. ^[28]

A related but distinct lineage developed in parallel at the University of Pennsylvania's Kod*Lab. Gavin Kenneally, Avik De, and Daniel Koditschek published "Design Principles for a Family of Direct-Drive Legged Robots" in 2016, describing the Minitaur quadruped (commercialized by Ghost Robotics), which eliminated the gearbox entirely rather than merely minimizing it, achieving direct-drive transparency, mechanical robustness, and high actuation bandwidth through a different route than MIT's low-ratio approach. ^[29] The term "quasi-direct drive" itself came into wide, explicit use over the following years; Stanford's 2019 paper "Stanford Doggo: An Open-Source, Quasi-Direct-Drive Quadruped," by Nathan Kau, Aaron Schultz, Natalie Ferrante, and Patrick Slade, used the phrase directly in its title and built on MIT's actuator design principles to create a low-cost, replicable quadruped for hobbyists and researchers. ^[21] By the early-to-mid 2020s the same design logic had propagated into commercial legged robots and into humanoid legs and hands, discussed below.

Use in legged robots and humanoids

Quadrupeds

QDD is now the default actuator style for most mass-produced quadruped robots. Unitree's Go2, B-series, and A2 platforms use single-stage planetary QDD actuators with reduction ratios below 10:1, and because actuators typically represent 50 to 70 percent of a quadruped's bill of materials, the lower cost of off-the-shelf planetary gearing (versus custom-machined harmonic flexsplines) has been a significant factor in the steep price declines seen across the category. ^[24] The tradeoff is durability: planetary gears used in QDD actuators wear and develop backlash faster than a comparable harmonic drive, so manufacturers increasingly treat the gearbox as a scheduled-replacement wear part rather than a component engineered to last a decade. ^[24]

Humanoid legs

Legs repeatedly slam a foot into the ground, and a backdrivable, low-reduction joint can absorb that impulse through the motor's own rotor and control loop rather than transmitting a shock spike through gear teeth, which reduces wear and supports more dynamic gaits such as running, jumping, and recovering from a stumble. ^{[1] [7] [20]} Unitree's G1 humanoid applies this logic with its roughly 20.6:1 two-stage planetary joints noted above. ^[16] Other humanoid programs have made the opposite bet for at least part of the body: Tesla's Optimus, as noted, pairs harmonic drives with dedicated torque sensors in its rotary joints for positioning precision, and Boston Dynamics' current all-electric Atlas uses custom rotary and linear actuators whose internal reduction architecture the company has not fully disclosed publicly. ^{[23] [30]} The split reflects genuine engineering disagreement about where a humanoid needs raw torque density and precision versus where it needs compliance and impact tolerance, and several companies are still converging on an answer joint by joint rather than committing one architecture to the whole body.

Hands and manipulation

The same transparency argument extends to dexterous hands, where space for sensors is especially tight. Early QDD manipulators, such as the low-cost compliant arm and gripper work published by Simon Kalouche in 2019, argued that QDD's inherent force sensing could replace costly load cells in manipulation as well as locomotion. ^[31] In July 2026, 1X Technologies unveiled a new hand for its 1X Neo humanoid that combines a tendon-driven transmission with quasi-direct-drive-style low gear ratios of roughly 5:1 to 15:1, delivering 25 total degrees of freedom (22 in the fingers and palm plus 3 at the wrist), peak torque of about 3.5 Nm at the thumb and 2.6 Nm at the fingers, and integrated tactile sensing at the fingertips, all demonstrated on tasks such as assembling small parts, using a screwdriver, and zipping a jacket. ^[32] On the research side, MIT's EyeSight Hand, developed by Branden Romero, Hao-Shu Fang, Pulkit Agrawal, and Edward Adelson, is a low-cost, 7-degrees-of-freedom dexterous hand built around QDD actuation paired with vision-based tactile sensors, intended to withstand the repeated, sometimes clumsy contact generated when collecting data for imitation-learning research. ^[33] In both cases the goal mirrors the leg case: let the hand feel contact forces through motor current rather than through a separate sensor at every joint, in a package too small and too cost-sensitive for one.

Reinforcement learning and sim-to-real transfer

Modern legged locomotion is increasingly trained with reinforcement learning in physics simulators, such as NVIDIA's Isaac Lab or MuJoCo, and then transferred to physical hardware, a process called sim-to-real transfer. QDD's comparatively simple, well-characterized dynamics, low friction, low backlash, and a fairly linear relationship between motor current and joint torque, are much easier to model accurately in simulation than the nonlinear friction, compliance, and stiction of a high-reduction gearbox, which narrows the gap between simulated and real-world behavior that often causes learned policies to fail on hardware. ^{[5] [11]} A widely cited example is "RMA: Rapid Motor Adaptation for Legged Robots" by Ankur Kumar (Ashish Kumar), Zipeng Fu, Deepak Pathak, and Jitendra Malik, which trained a locomotion policy entirely in simulation and deployed it, without any real-world fine-tuning, on Unitree's QDD-actuated A1 quadruped, successfully crossing sand, mud, and obstacle-strewn terrain. ^[34] Where a design departs from a clean single-stage planetary QDD layout, such as the cycloidal QDD actuators explored for legged robots, researchers have built learned "actuator network" models specifically to correct for the resulting nonlinearity and preserve the sim-to-real benefit. ^[5] Backdrivability also carries a safety dimension: because a QDD joint yields naturally under unexpected contact rather than fighting it, it is inherently gentler in accidental collisions with people or the environment than a stiff, high-reduction joint would be. ^[6]

Notable QDD platforms and actuators

Scroll sideways for more →

See also

• Actuator
• Harmonic drive
• Cycloidal drive
• Planetary gear train
• Brushless DC motor
• Force/torque sensor
• Reinforcement learning
• Sim-to-real transfer
• Unitree
• Humanoid robot
• Dexterous hand
• Tendon-driven

References

1. Wensing, P.M., Wang, A., Seok, S., Otten, D., Lang, J., Kim, S. "Proprioceptive Actuator Design in the MIT Cheetah: Impact Mitigation and High-Bandwidth Physical Interaction for Dynamic Legged Robots." IEEE Transactions on Robotics, 2017. https://fab.cba.mit.edu/classes/865.18/motion/papers/mit-cheetah-actuator.pdf ↩
2. Emergent Mind. "Quasi-Direct-Drive (QDD) Actuation." https://www.emergentmind.com/topics/quasi-direct-drive-qdd-actuation ↩
3. ZHR Motor / Robotics Zhinno. "Harmonic Reducer vs Planetary Reducer: Data-Driven Comparison for Robot Engineers." https://robotics.zhinno.com/blog/harmonic-reducer-vs-planetary-reducer.html ↩
4. EYOU Robotics. "Types of Robot Joint Actuators: Harmonic, Planetary, QDD & More." https://eyoubot.com/en/blog/types-of-robot-joint-actuators ↩
5. Zhu, A., Tanaka, Y., Rafeedi, F., Hong, D. "Cycloidal Quasi-Direct Drive Actuator Designs with Learning-based Torque Estimation for Legged Robotics." IEEE International Conference on Robotics and Automation, 2025. https://arxiv.org/abs/2410.16591 ↩
6. ResearchGate (multiple authors). "Factors Influencing Actuator's Backdrivability in Human-Centered Robotics." https://www.researchgate.net/publication/364208810_Factors_influencing_actuator's_backdrivability_in_human-centered_robotics ↩
7. SemiAnalysis. "Quadruped State of The Market: Unitree, Boston Dynamics, ANYbotics, DEEP Robotics, and The Rising Application Ecosystem." https://newsletter.semianalysis.com/p/quadruped-state-of-the-market-unitree ↩
8. Urs, K., Enninful Adu, C., Rouse, E.J., Moore, T.Y. "Alternative Metrics to Select Motors for Quasi-Direct Drive Actuators." 2022. https://arxiv.org/pdf/2202.12365 ↩
9. Jeong, S. "A Cycloidal Quasi-Direct Drive Actuator: Design, Fabrication and Control." Personal engineering blog. https://jeongseojin.github.io/blog/2025/qdd-actuator/ ↩
10. Brushless.com. "How to Calculate the KV Value of Brushless Motor." https://www.brushless.com/how-to-calculate-the-kv-value-of-brushless-motor ↩
11. Singh, A., Kapa, D., Chedda, P., Kolathaya, S.N.Y. "Comparison between External and Internal Single Stage Planetary Gearbox Actuators for Legged Robots." Advances in Robotics, 2025. https://arxiv.org/abs/2506.16356 ↩
12. GitHub (community project documentation). "Quasi-Direct-Drive Actuator." https://github.com/JeongSeoJin/Quasi-Direct-Drive-Actuator ↩
13. Singh, A., Kapa, D., Chedda, P., Kolathaya, S.N.Y. "Comparison between External and Internal Single Stage Planetary Gearbox Actuators for Legged Robots." Advances in Robotics, 2025. https://arxiv.org/html/2506.16356v1 ↩
14. Seok, S., Wang, A., Chuah, M.Y., Otten, D., Lang, J., Kim, S. "Design Principles for Energy-Efficient Legged Locomotion and Implementation on the MIT Cheetah Robot." IEEE/ASME Transactions on Mechatronics (extended from IEEE ICRA 2013). https://dspace.mit.edu/bitstream/handle/1721.1/108096/Efficiency-Principles-for-Quadrupeds_v3.pdf ↩
15. MIT Biomimetic Robotics Lab. "Research." https://biomimetics.mit.edu/research ↩
16. Robotopian. "Unitree G1 Humanoid Robot Teardown and Commercial Deployment Analysis." https://robotopian.com/blogs/news/unitree-g1-humanoid-robot-teardown ↩
17. HD Harmonic. "What Is the Gear Ratio of a Harmonic Drive? Exploring the Mechanics Behind Precision." https://www.hdharmonic.com/news/what-is-the-gear-ratio-of-a-harmonic-drive-exploring-the-mechanics-behind-precision ↩
18. Electronic Design. "Harmonic Drive Gear Has 30:1 Gear Ratio." https://www.electronicdesign.com/technologies/components/electromechanical/article/21748881/harmonic-drive-gear-has-301-gear-ratio ↩
19. Sumitomo Drive Technologies. "Cyclo Reducer." https://canada.sumitomodrive.com/en-ca/product/cyclo-reducer ↩
20. ZHR Motor / Robotics Zhinno. "Robot Actuator Torque Density Comparison: Harmonic vs Planetary." https://robotics.zhinno.com/blog/robot-actuator-torque-density-comparison.html ↩
21. Kau, N., Schultz, A., Ferrante, N., Slade, P. "Stanford Doggo: An Open-Source, Quasi-Direct-Drive Quadruped." IEEE International Conference on Robotics and Automation, 2019. https://arxiv.org/pdf/1905.04254 ↩
22. LinkedIn (Inspire Robots). "A Technical Breakdown of Tesla Optimus' Linear Actuators." https://www.linkedin.com/pulse/inspire-robots-technical-breakdown-tesla-optimus-linear--1c ↩
23. Optimusk.blog. "Tesla Optimus Hardware: Actuators, Hands & Sensors (2026)." https://optimusk.blog/blog/tesla-optimus-hardware-specs/ ↩
24. SemiAnalysis. "Quadruped State of The Market: Unitree, Boston Dynamics, ANYbotics, DEEP Robotics, and The Rising Application Ecosystem." https://newsletter.semianalysis.com/p/quadruped-state-of-the-market-unitree ↩
25. Robots Guide. "Mini Cheetah." https://robotsguide.com/robots/minicheetah ↩
26. MIT Open Access Articles. "MIT Cheetah 3: Design and Control of a Robust, Dynamic Quadruped Robot." IEEE/RSJ International Conference on Intelligent Robots and Systems, 2018. https://dspace.mit.edu/bitstream/handle/1721.1/126619/IROS.pdf ↩
27. MIT News. "Mini Cheetah Is the First Four-Legged Robot to Do a Backflip." March 4, 2019. https://news.mit.edu/2019/mit-mini-cheetah-first-four-legged-robot-to-backflip-0304 ↩
28. IEEE Spectrum. "MIT Is Building a Dynamic, Acrobatic Humanoid Robot." https://spectrum.ieee.org/mit-dynamic-acrobatic-humanoid-robot ↩
29. Kenneally, G., De, A., Koditschek, D.E. "Design Principles for a Family of Direct-Drive Legged Robots." IEEE Robotics and Automation Letters, 2016. https://kodlab.seas.upenn.edu/uploads/Gavin/gake_dddesign.pdf ↩
30. AI2Work. "Boston Dynamics Electric Atlas Ships Its First Commercial Units." https://ai2.work/blog/boston-dynamics-electric-atlas-ships-its-first-commercial-units ↩
31. Kalouche, S. "Quasi-Direct Drive for Low-Cost Compliant Robotic Manipulation." 2019. https://arxiv.org/pdf/1904.03815 ↩
32. The AI Insider. "1X Unveils New Dexterous Tendon-driven Hand for Its Neo Humanoid Robots." July 9, 2026. https://theaiinsider.tech/2026/07/09/1x-unveils-new-dexterous-tendon-driven-hand-for-its-neo-humanoid-robots/ ↩
33. Romero, B., Fang, H.S., Agrawal, P., Adelson, E. "EyeSight Hand: Design of a Fully-Actuated Dexterous Robot Hand with Integrated Vision-Based Tactile Sensors and Compliant Actuation." 2024. https://arxiv.org/pdf/2408.06265 ↩
34. Kumar, A., Fu, Z., Pathak, D., Malik, J. "RMA: Rapid Motor Adaptation for Legged Robots." Robotics: Science and Systems, 2021. https://arxiv.org/abs/2107.04034 ↩