I have a rolling ball as an agent which receives a force applied during on action received. I would like to use a ray perception sensor but it rotates with the ball. I’ve added this sensor as a child and during late update I reset the rotation, which works fine when testing in the editor. However, when I attempt to train it still rotates the sensor somewhat if I have the time scale set to a value above 1.
Is there a general rule of thumb when it comes to physics simulations and time scale?
[Edit] Vincentpierre gives a better explanation below for the time scale.
You can lock the x,y,z rotation on your ball’s rigidbody and just use a physics material to control roll-like-behavior via the friction. Or if you really need the ball to actually roll you can just put a rigidbody on the ray perception sensor, lock it’s rotation, and just update it’s rigidbody.position in FixedUpate() to match the ball.
LateUpdate is on the same cycle as Update. FixedUpdate is on a different cycle. ML-Agents methods are called in the FixedUpdate cycle, so it is not recommended to have some logic of the same Agent be split over FixedUpdate and LateUpdate.
If you increase the time scale, the ratio of Updates vs FixedUpdates will be changed and this is why you see the rotation more when using a larger timescale.
Thanks for the help. FixedUpdate did the trick.