Miguel Alonso JrI've fallen <from a random position and orientation with all my joints at random angles> and I... can get up? #mlagents #reinforcementlearning #machinelearning #artificialintelligence
marcorobotics@miguelalonsojr the standing poses look a bit unnatural. Is the center of mass at the front of the robot or closer to the backpack?
@marcorobotics I haven't taken a look at where the center of mass is. I used this Atlas model (https://github.com/mrmarten/atlas_ros_version). This policy has not been fully trained... it's about 1/2 way there. That's why it still looks funny. I'll probably have to shape the reward a bit more to get a more natural motion, but this is actually pretty good so far. I'll post a vid of the fully trained policy once it's complete.
@marcorobotics I also suspect I'll have to make more adjustments to the Atlas as I'm just eyeballing things to get something stood up by December. It will definitely need to be tuned further.
@marcorobotics I've also been discovering some problems with the URDF import and have had to adjust. I just found another one in the ankle joints. Things like incorrect joint limits, colliders being off, etc. Eventually I'll work out all of the bugs.