System Identification

Using measured input-output data to work backward to a system's mathematical model and parameters.

System identification is the discipline of building a mathematical model of a dynamic system from measured input and output data using statistical methods, and it also covers how to design experiments that gather sufficiently informative data. It is grouped into three categories by how much prior knowledge is used: white-box (derived entirely from physical laws), gray-box (known structure with parameters fit from data, e.g., identifying link mass, inertia, and joint friction), and black-box (fitting only the input-output relationship, e.g., with a neural network). In embodied AI, it is one of the main tools for narrowing the sim-to-real gap: parameters measured on the real robot are written back into the simulator, or a model is learned to fill in whatever the simulator is missing. Hwangbo et al. (2019) first identified ANYmal's physical parameters, then used real-robot data to train an actuator network modeling the motor and low-level software dynamics. It complements domain randomization: identification makes the simulation more accurate, while randomization makes the policy robust to whatever error remains.

ExampleApply a sinusoidal sweep of torque to one robot joint, record its angle and angular velocity, and fit the moment of inertia, viscous friction, and Coulomb friction coefficients by least squares; writing these back into a MuJoCo model aligns the simulated joint trajectory with the real robot under the same commands.

Also called

SysID, Parameter Identification

Related

Dynamic Parameter Identification · Actuator Modeling (Actuator Network) · Sim-to-Real Gap (Reality Gap) · Domain Randomization · Real-to-Sim · Inertial Parameters

Sources

Wikipedia: System identification
Learning agile and dynamic motor skills for legged robots（执行器网络） (Chinese)