Robots Atlas>ROBOTS ATLAS
Robotics

Back-drivable actuation

ActivePublished: 31 July 2026Updated: 23 September 2026Published
Key innovation
Back-drivable actuation: an external force applied to a joint's output can turn the motor backwards, thanks to low reflected inertia and low friction. It provides mechanical transparency, force/torque control and compliance without dedicated force sensors.
Category
Robotics
Abstraction level
Building block
Operation level
Robot control
Use cases
Safe human-robot interaction (cobots, humanoids)Force/torque control and compliance (impedance control)Kinesthetic teaching and teleoperation (collecting demonstrations)Dynamic locomotion and impact absorption (legged robots)Delicate manipulation and contact sensing

How it works

1) The motor couples to the output through a drivetrain with low reflected inertia and low friction (e.g. quasi-direct-drive, low gear ratio). 2) An external force at the output easily turns the motor backwards (transparency). 3) Output torque is estimated from motor current/torque (proprioceptive actuator), without a separate force sensor. 4) The controller implements torque/impedance control (compliance), enabling safe contact, impact absorption and kinesthetic guiding of the arm. Back-drivability improves when the motor rotor's reflected inertia is low - and that inertia scales with the square of the gear ratio (~N^2 x J_motor), which is why high gear ratios sharply degrade it.

Problem solved

Self-locking drives (high gear ratios) are stiff, opaque and unsafe on contact and hard to force-control. Back-drivable actuation provides compliance, torque control and safe physical interaction, often without a dedicated force sensor.

Components

High-torque motorTorque source

Typically a BLDC/PMSM motor capable of torque control without a high gear ratio.

Official

Low reflected-inertia drivetrainMechanical transparency

Low gear ratio and an efficient transmission minimizing reflected inertia and friction (back-drivability).

Official

Current-based torque estimationForce sensing

Sensing output torque from motor current, without a dedicated force sensor.

Official

Torque/impedance controlCompliance and force

A control loop implementing compliance and the desired output torque.

Official

Implementation

Implementation pitfalls
Torque vs back-drivability trade-offHigh

A low gear ratio improves back-drivability but requires a larger/heavier motor for a given torque.

Fix:Quasi-direct-drive with a high-torque motor, mass/torque optimization, or SEA.
Accuracy of current-based torque estimationMedium

Friction, torque ripple (cogging) and temperature introduce errors in sensorless force estimation.

Fix:Calibration, friction/cogging compensation, an optional torque sensor.
Holding position without powerMedium

A back-drivable drive is not self-locking - on power loss the joint may drop under load.

Fix:Safety brakes, fail-safe control, gravity compensation.

Evolution

1995
Series Elastic Actuators (SEA)

Pratt and Williamson introduce series elastic actuators - an early approach to compliant, force-controlled actuation.

2017
Proprioceptive actuators (MIT Cheetah)
Inflection point

Wensing et al. describe the proprioceptive actuator design in the MIT Cheetah - a highly back-drivable quasi-direct-drive for dynamic locomotion and physical interaction. The paper also introduces the IMF (impact mitigation factor) metric to quantify back-drivability at impact.

2023
Standard in humanoids and cobots

Back-drivable actuators (often quasi-direct-drive) become standard in modern humanoids and cobots for safety and data collection.

Hyperparameters (configurable axes)

Gear ratioCritical

A lower gear ratio means more back-drivability and transparency but lower output torque. The motor's reflected inertia scales with the square of the gear ratio (~N^2).

Transmission typeHigh

Quasi-direct-drive/planetary (back-drivable) vs worm/harmonic with high friction (less back-drivable).

Torque sensing methodMedium

Current-based estimation (proprioceptive) vs a dedicated torque sensor.

Computational complexity

Time complexity: Nie dotyczy (wlasciwosc mechatroniczna napedu, nie algorytm). Space complexity: Nie dotyczy.

Compute bottleneck

Real-time torque control loop

Back-drivability itself is a mechanical property; what matters computationally is a fast (kHz) torque/impedance control loop and current-based torque estimation.

Depends on
Czestotliwosc petli sterowaniaDokladnosc estymacji momentu

Execution paradigm

Primary mode
Conditional

This describes the actuator control loop; back-drivability itself is a mechanical property, not a compute paradigm.

Activation pattern
Input dependent
Routing mechanism

Parallelism

Parallelism level
Fully parallel

Each actuator/joint has its own torque control loop, easy to parallelize in hardware.

Scope
Inference
Constraints
!Per-joint torque loops can run in parallel (though whole-body control ties them together).

Hardware requirements

Primary

A mechatronic actuator property, independent of the compute accelerator type; the control loop typically runs on the drive's microcontroller/DSP.