Dynamic Parameter Identification

动力学参数辨识Advanced

Moving the robot through specific trajectories and using the torque data to work out each link's mass and inertia.

Dynamic parameter identification means estimating the parameters of a robot's dynamics model from measured data: each link's mass, first moment of mass (related to the center-of-mass location), and inertia tensor — 10 inertial parameters per link in total, often together with joint friction. Atkeson, An, and Hollerbach showed in 1986 that the Newton-Euler equations can be rewritten so joint torque is linear in these parameters, τ = Y(q,q̇,q̈)π, where Y is a regressor matrix depending only on joint position, velocity, and acceleration, and π is the parameter vector — making least squares sufficient to solve for them. To excite every parameter well enough in the data, the arm is often driven along an excitation trajectory optimized as a Fourier series (Swevers et al., 1997). CAD-supplied parameters are often inaccurate, and they change further once a payload is attached; accurate identification is what makes gravity compensation, computed-torque control, and sim-to-real alignment reliable. Wensing and colleagues added linear-matrix-inequality constraints in 2017 to guarantee the identified parameters remain physically consistent.

ExampleAfter a new gripper with unknown parameters is mounted on an arm's end effector, the joints are driven through a periodic motion for a few dozen seconds while joint angles and torques (or currents) are logged; least squares then solves for the payload's mass and center of mass, which are fed back into the gravity-compensation model.

Also called: Inertial Parameter Identification, Payload Identification
Related: Inertial Parameters · System Identification · Rigid-Body Dynamics · Gravity Compensation · Friction Compensation · Payload
Sources: Atkeson, An, Hollerbach: Estimation of Inertial Parameters of Manipulator Loads and Links (IJRR 1986)
Swevers et al.: Optimal robot excitation and identification (IEEE TRA 1997)
Wensing, Kim, Slotine: Linear Matrix Inequalities for Physically-Consistent Inertial Parameter Identification