Open axnedergaard opened 2 years ago
This is due to the checkpoints being stored at the end of an episode and the episode in jaco is (sometimes?) only limited to 249. This is even though the time limit is 10s
and the time per step is 0.04s
which should result in 250 steps, but I guess the limit is only sometimes included. So the solution for this would be either to store the checkpoint at each step at the specified time steps or to increase the time limit to 10.04s
.
I ran pretrain.py for various agents (rnd, icm, proto, icm_apt) and domains (walker, quadruped, jaco).
After the code ran, the directory models/states only contains the subdirectories walker and quadruped.
To reproduce:
python pretrain.py agent=rnd domain=jaco
orpython pretrain.py agent=icm domain=jaco
orpython pretrain.py agent=proto domain=jaco
orpython pretrain.py agent=icm_apt domain=jaco