OpenAI Gym Wrapper for the DeepMind Control Suite
Note that there's some issue on Nvidia OpenGL. Rendering backend is specified to the OSMesa, which can be modified on the head of dm2gym.py
.
RuntimeError
will be raised after the program has been terminated, which is by design.
See issue in dm_control for details.
Except for the APIs listed as follows, others are consistent with OpenAI Gym.
make
is consistent withsuite.load
indm_control
.env.render
has four parameters where one can specify cameras or mode. More details can be found in codes.
Simple usages:
import dm2gym
env = dm2gym.make(domain_name="cartpole", task_name="balance")
env.reset()
for t in range(1000):
env.render('human', camera_ids=[0, 1])
a = env.action_space.sample()
observation, reward, done, info = env.step(a)
if done:
env.reset()