Question My robot arm falls apart and gets stuck while training in ml-agents/

Hi all,

I'm training my robot arm these days. After I press the play button, the arm itself can move fluently and it seems that there isn't any question.

However, when I use my .yaml to train it and then press the play button when the terminal calls me to do it, my robot immediately falls apart and doesn't move.

I wonder why it doesn't move properly while training but falls apart. Is there a problem with my arm or the wrong .yaml file will cause this thing to happen?

 

How is the arm setup - if its an articulation body then check all your drive limits and increase your solver frequency.
How is your agent driving the arm? If its randomly choosing a drive limit to apply then you could look into driving the arm incrementally (more like ROS does)

 

Make sure that you can drive each individual joint through the heuristic function; you want to emulate the same types of inputs that the NN model is sending to the OnActionRecieved -> your arm. For things like this you can create a function that simply oscillates between -1 and 1 and use that value in the heuristic function to drive joints, then control it from the editor using a Boolean.

If the arm works through heuristic mode, but not in training (through the yaml config), then something else is going on.

 

Yeah, thanks for both of you quite a lot!!

I previously use the configurable joint and don't apply parent child relationship to the robot arm, which seems to be very unstable. And now I apply parent child relationship and articulation body instead, which becomes stable!

Although at first when I still used configurable joint, the robot arm may start moving properly at first after starting training and falling apart after a few seconds, which really confused me.

Again, thanks for help!