Actuator Modeling (Actuator Network)

Simulating how a real motor's torque output actually responds to a command, to narrow the sim-to-real gap.

Actuator modeling means describing, inside a simulation, how a real motor and gearbox actually behave: given a target joint position, how much torque it can really deliver, and with how much delay and saturation. A simulator's default ideal PD control responds faster and more cleanly than a real motor does, and a policy trained against that ideal often jitters or lags once it reaches a real robot — a significant source of the sim-to-real gap. There are two general approaches: analytical motor models that add torque limits, speed saturation, or command delay; and actuator networks, which originated in ETH Zurich and Intel's 2019 ANYmal work, where a small neural network is trained on less than 4 minutes of real-robot data to predict torque from a short history of joint position error and velocity, and is then plugged into the simulator for policy training. Isaac Lab calls the case where the physics engine computes PD control internally an “implicit actuator,” and calls this kind of custom model an “explicit actuator.”

ExampleIsaac Lab provides actuator classes such as DCMotor (with speed-torque saturation), DelayedPDActuator (with command delay), and the neural-network-based ActuatorNetMLP and ActuatorNetLSTM, chosen to match the real motor's characteristics when training a quadruped policy.

Also called

Actuator Network, Actuator Net, implicit actuator, explicit actuator

Related

Sim-to-Real Transfer · Sim-to-Real Gap (Reality Gap) · System Identification · Proportional-Derivative Control · Series Elastic Actuator (SEA) · NVIDIA Isaac Lab

Sources

Learning agile and dynamic motor skills for legged robots (Hwangbo et al., Science Robotics 2019)
Isaac Lab Docs: Actuators
Isaac Lab API: isaaclab.actuators

As of

2026-09