go1 learns 3 legged policy
Opened this issue · 3 comments
Hello, when I try training the base policy for go2 python train.py --exptid 054-01-go1 --task go1 it learns to only use the front 2 legs and the back left leg. I tried training the a1 policy to see if it happens but that turned out fine and then I tried training go1 again and it learned 3 legs again. Is this something with the go1 config?
Go1 support is not added yet, so training with --task go1 will have unexpected behaviors especially the terrains contain extreme cases since the max torque for go1 is 23.7Nm vs A1~33 Nm. You can try to change the urdf here instead of using --task go1 and see what happens.
I am also interested in training the go1 policy. May I ask that have you successfully deployed this algorithm after modifying the urdf? By the way, the go1 urdf file in Isaac Sim seems to be the old version.
some problem with you. When I try to train the base policy for other quals, it always learns to use the front leg and the back left leg. I tried to lower the difficulty of the terrain and reduce the rew_feet_edge, but they don't work. May I ask you that have you successfully deploy the alogorithm or fix the problem?