Dynamics Randomization

Randomizing physical parameters like mass and friction during training so a policy can handle a real robot.

Dynamics randomization is a form of domain randomization that specifically randomizes physical parameters rather than visual appearance. A simulation's mass, friction, damping, motor gains, and delay can never fully match a real robot, and training under one fixed set of these values leaves a policy prone to overfitting to the simulator. The approach is to resample these parameters from a chosen range at the start of every episode, forcing the policy to succeed under a wide variety of physical conditions; if the real robot's true parameters fall within that range, the policy is more likely to work on it directly. A landmark example is Peng and colleagues' 2018 paper, which randomized 95 parameters for a puck-pushing task and deployed a policy trained purely in simulation directly onto a Fetch robot arm, with performance close to simulation. Today, frameworks such as Isaac Lab and legged_gym routinely randomize ground friction and payload mass, and apply random pushes, when training legged robots.

Examplelegged_gym by default samples ground friction uniformly between 0.5 and 1.25, and randomly shoves the robot every 15 seconds, forcing the policy to learn to keep walking across different surfaces and under unexpected disturbances.

Also called

physics parameter randomization, friction/mass randomization

Related

Domain Randomization · Sim-to-Real Transfer · Sim-to-Real Gap (Reality Gap) · Visual Randomization · Automatic Domain Randomization · System Identification

Sources

Sim-to-Real Transfer of Robotic Control with Dynamics Randomization (arXiv 1710.06537)
legged_gym: legged_robot_config.py（domain_rand）