RecordEpisode() environment wrapper for 'TwoRobotPickCube-v1' task

[seungbinjoo]

Hi, I'm a user looking to use ManiSkill for RL for bimanual manipulation. I'm just testing recording episodes when executing a random policy for the 'TwoRobotPickCube-v1' task like so (following the quickstart notebook):

# to make it look a little more realistic, we will enable shadows which make the default lighting cast shadows
env = gym.make("TwoRobotPickCube-v1", num_envs=4, render_mode="rgb_array", enable_shadow=True)
env = RecordEpisode(
    env,
    "./videos", # the directory to save replay videos and trajectories to
    # on GPU sim we record intervals, not by single episodes as there are multiple envs
    # each 100 steps a new video is saved
    max_steps_per_video=100
)

# step through the environment with random actions
obs, _ = env.reset()
for i in range(100):
    action = env.action_space.sample()
    obs, reward, terminated, truncated, info = env.step(action)
    # env.render_human() # will render with a window if possible
env.close()
from IPython.display import Video
Video("./videos/0.mp4", embed=True, width=640) # Watch the replay

But I run into the following issue:

Traceback (most recent call last):
  File "/athenahomes/joo/4yp/rma4rma/troubleshooting/test_mani_skill3.py", line 61, in <module>
    obs, _ = env.reset()
             ^^^^^^^^^^^
  File "/users/joo/miniconda3/envs/4YP/lib/python3.11/site-packages/mani_skill/utils/wrappers/record.py", line 367, in reset
    action=common.to_numpy(common.batch(action.repeat(self.num_envs, 0))),
                                        ^^^^^^^^^^^^^
AttributeError: 'dict' object has no attribute 'repeat'

[StoneT2000]

hm this is due to the default dictionary action space.

I've seen multiple issues around the fact the action space typically needs to be flattened before use for most algorithms / workflows that I am considering making flat the default.

The dictionary action space was originally designed to support multi agent RL setups.

For a workaround for now you can first apply the FlattenActionSpace wrapper: https://maniskill.readthedocs.io/en/latest/user_guide/wrappers/flatten.html