Model Identification

Derivative-Free Online Learning of Inverse Dynamics Models

“…k = 5) have also been tested leading to similar results and have been therefore omitted. The results in Figure 6 show that the performance of NP-ML-DF (which uses R = I) is 8 The transient behaviour has been omitted because it does not add much information with respect to the steady state statistics. worse than NP-ML-DFR with k = 3.…”

“…Since the RBD model is physics-based, it should, in principle, describe the robot dynamics for all the desired trajectories, leading to good generalization and global approximation properties. A known issue of this model (see e.g., [8]) is that the problem of determining π from measured data y(t) is usually ill posed and the matrix ψ(x(t)) is rank deficient. Possible solutions in system identification are either the design of efficient experiments to collect data sufficiently rich to excite the highest number of modes of the system or dedicated experiments which are good to estimate parameters separately.…”

Robot Self-Calibration Using Multiple Kinematic Chains—A Simulation Study on the iCub Humanoid Robot

“…Most often, the loop is closed through self-observation of the end-effector using cameras located in the robot head (open-loop calibration method per [1]). Hersch et al [7] and Martinez-Cantin et al [8] present online methods to calibrate humanoid torso kinematics relying on gradient descent and recursive least squares estimation, respectively.…”

“…The robot kinematics is represented using the standard Denavit-Hartenberg (DH) parameters. Hollerbach et al [1] classify different calibration methods into openloop-where one or more of the components of the endeffector pose is measured employing mechanical, visual, or laser systems-and closed-loop where physical constraints on the end-effector position or orientation can substitute for measurements. 1 Observing the end-effector-or in general any other points on the kinematic chain-using a camera falls into the open-loop calibration family, even though components of the end-effectors pose can be observed only indirectly through projection into the camera frame.…”

“…Hollerbach et al [1] classify different calibration methods into openloop-where one or more of the components of the endeffector pose is measured employing mechanical, visual, or laser systems-and closed-loop where physical constraints on the end-effector position or orientation can substitute for measurements. 1 Observing the end-effector-or in general any other points on the kinematic chain-using a camera falls into the open-loop calibration family, even though components of the end-effectors pose can be observed only indirectly through projection into the camera frame. The selftouch configurations-in this work end-effectors of the two arms in contact-can in our case be regarded as open-loop as well, since the contact does not act as a physical constraint but rather as measurement: a tactile or force sensor would signal that the two end-effectors are in the same position (i.e.…”