I have no idea what’s causing this issue, and can’t figure out what’s happening. As certain points during the training, the VectorAction will suddenly spit out NaN values. Sometimes, it’ll dissapear in a few seconds, but some other times, it doesn’t go away, and I have to start a new training run. Any ideas?
Did you check for divisions by zero?
EDIT Oops, you’re talking about actions, not observations. Not sure if I remember correctly, but I think I’ve seen NaN actions following NaN observations before. Can’t hurt to check for NaN observations anyway.
All my observations are direct measurements from the rigidbodys(velocity, position, rotation). I’m not dividing any of them
Hi MapleMeese. What version of ML-Agents are you using? This may have been a bug which has been resolved in a newer release.
I’m on the latest version in the package manager V 1.0.6