[Robot Hardware 04] - Actuators (3): QDD Actuators

Robot hardware from a Physical AI perspective - QDD actuators

EN KR

Robot Hardware from a Physical AI perspective

The high-ratio systems discussed in the previous notes, such as Harmonic Drives, use a small motor and a high-ratio reducer to obtain torque. A QDD (Quasi-Direct Drive) actuator takes almost the opposite design philosophy.

The main goals of QDD are:

  1. Low Mechanical Impedance: Keep the mechanical effect of motor rotor inertia low.
  2. High Transparency: Minimize mechanical nonlinearities such as friction and backlash.
  3. High Bandwidth: Respond quickly enough to follow high-frequency force and motion commands.

These properties are useful in systems that interact dynamically with the environment, such as quadrupeds and humanoids.

Structure and Design

QDD is not usually a fully direct-drive system. Instead, it uses a low reduction ratio, commonly around 6:1 to 10:1.

Because the reducer no longer provides a large torque multiplier, QDD actuators rely on high-torque-density motors.

  • Motor: Large-diameter, high-pole-count BLDC outrunner motors are common because a larger air-gap radius helps produce more torque.
  • Reducer: Planetary gears or cycloidal drives are often used. To keep the package compact, the reducer may be placed inside the open space of the motor stator.
Example QDD actuator [1]
MIT Mini Cheetah Actuator Teardown

This structure makes it possible to build a compact joint module where the motor and reducer are tightly integrated.

Core Concept: Force-Motion Transparency

The central performance concept for QDD is force-motion transparency.

This term originally comes from teleoperation. It describes how accurately force and motion at the input side are transmitted to the output side, and how cleanly reaction forces at the output side are transmitted back to the input side.

Using a mechanical impedance model $Z(s)$, high transparency means that internal inertia $M$ and friction $B$ should be as small as possible.

\[Z(s) = \frac{F(s)}{v(s)} = Ms + B + \frac{K}{s}\]

By lowering the reduction ratio, QDD reduces both high-frequency distortion from inertia and low-frequency distortion from friction. The goal is to preserve transparency across a wide frequency range.

Two Ways to Define QDD

There is no single strict numerical definition of QDD. In practice, it is usually discussed from two perspectives.

1. Dynamic Transparency

This perspective focuses on impact mitigation and high-speed control. The key quantity is reflected inertia.

Motor rotor inertia $J_m$ is amplified through a reducer by the square of the gear ratio $N^2$:

\[J_{ref} = J_m \times N^2\]

In a high-ratio system, such as 100:1, rotor inertia is amplified by 10,000 times. The joint becomes dynamically sluggish. In a QDD system, such as 6:1, the amplification is only 36 times.

Benefits of low reflected inertia:

  • Impact mitigation: The actuator can mechanically absorb or pass through impact energy during ground contact or collision.
  • High control bandwidth: Low inertia allows faster acceleration and better tracking of high-frequency commands.

2. Quasi-Static Force Transparency

This perspective focuses on how accurately force can be estimated from motor current.

Even if the gear ratio is moderately high, around 20:1 for example, an actuator may still be considered QDD-like if mechanical nonlinearities such as friction and hysteresis are sufficiently low.

  • High backdrivability: The output shaft should move smoothly under external force, even when power is off.
  • Linearity: Low friction helps preserve a linear relationship between current $I$ and torque $\tau$.

Proprioceptive Actuation

One control idea enabled by the force transparency of QDD hardware is proprioceptive actuation.

This means estimating and controlling external force using only motor current, without a separate torque sensor.

In a QDD system:

  1. Gear friction and mechanical nonlinearities are low.
  2. Reflected inertia is small, so dynamic disturbance is reduced.
  3. The relationship motor current $\propto$ external torque becomes usable.

This allows force sensing and force control with a simpler sensor configuration.

Note: This idea was introduced in the MIT Biomimetic Robotics Lab paper “Proprioceptive Actuator Design in the MIT Cheetah” [2]. I will discuss it in more detail later.

Synergy with Learning-Based Control and Its Limits

Fit with Learning-Based Control

QDD offers hardware properties that are useful when deploying learning-based controllers, especially policies trained with reinforcement learning, on real robots.

This is not because QDD changes the learning algorithm itself. Rather, it helps the real actuator behave more like the actuator assumed by the policy.

  • Reduced sim-to-real gap: High-ratio transmissions often introduce large friction, backlash, stiction, and complex transmission nonlinearities. QDD reduces many of these effects, so the gap between simulated actuator dynamics and real hardware response becomes smaller. This can reduce performance loss when transferring policies from simulation to hardware.

  • Backdrivability: Because QDD has high mechanical transparency, the actuator does not strongly reject external contact or unexpected collision. This is useful in manipulation and locomotion, where frequent contact makes overly rigid behavior unstable.

Limitation: Heat

The main weakness of QDD is heat.

Because the reduction ratio is low, the motor itself must generate much of the required torque. That means high current. Copper loss in the motor, $P_{loss} = I^2R$, grows with the square of current, so heat rises quickly.

  • Low torque multiplication: With less help from the gear ratio, the actuator depends heavily on the motor’s own torque constant.
  • Thermal management: Static tasks that require the robot to hold high gravitational loads can be inefficient and may overheat the motor.

For this reason, QDD actuators are often better suited to dynamic tasks where the robot is continuously moving, exchanging energy, and using regenerative effects. In real systems, motor thermal modeling and cooling design are essential.

References

[1] https://www.semanticscholar.org/paper/A-low-cost-modular-actuator-for-dynamic-robots-Katz/80732f8a46655aa4a1037a7fbdc154f4ceb33c50

[2] https://fab.cba.mit.edu/classes/865.18/motion/papers/mit-cheetah-actuator.pdf