How to inject noise into the supervisor’s policy while demonstrating? about dart HOT 1 CLOSED

Hi! One way to do it on a continuous control robot is to start with the control that the supervisor intends to send to the robot in the form of a vector, for example u = [1, 1], and then get a random sample from a normal/gaussian distribution, for example v = np.random.multivariate_normal(np.zerso(2), cov), where cov is the covariance matrix computed from DART or something you choose. Then instead of sending u to the robot, you can send u + v, which makes u noisy. You can think of it as intercepting the supervisor's intended control, modifying it, and then passing the modified version on to the robot.

from dart.