The pixel-based dmc2gym environment cannot run.

[LGXHTM]

How can I run the dmc2gym program? I successfully executed pip install gym and pip install dm_control in the environment, but the command pip install git+git://github.com/denisyarats/dmc2gym.git failed. I then downloaded the dmc2gym source code and successfully ran python setup.py install. When I executed dmc2gym_state_smz_config.py, it worked properly. However, when I executed dmc2gym_pixels_sez_config.py, an error occurred, which seems to be related to rendering and environment initialization. Additionally, I noticed that the model folder is missing a file named sampled_muzero_model.py. How can I resolve this issue? Thank you for your response.

[puyuan1996]

Thank you for your feedback! We will work on the fix in the coming days. If you have any other questions, feel free to let us know. We appreciate your support!

[puyuan1996]

Hello, regarding our previous research on DMC state inputs, you can refer to the relevant experimental results in https://arxiv.org/abs/2406.10667. Currently, I have fixed the relevant parts for DMC pixel in this PR (#322) and added the sampled_muzero_model. Moreover, I’ve locally tested LightZero/zoo/dmc2gym/config/dmc2gym_pixels_smz_config.py, and training works as expected. I will update the experimental curves in due course.

The package versions I used are as follows:

pip install gym==0.25.1 mujoco==3.2.6 dmc2gym==1.0.0

Please note that if you encounter errors related to DMC rendering, refer to the instructions here: https://github.com/google-deepmind/dm_control?tab=readme-ov-file#rendering. If your machine is in a headless environment, make sure to set the following environment variable before running the program:

export MUJOCO_GL=osmesa

Thank you for your support!

[puyuan1996]

Hello! Below are the collected and evaluation curves of Sampled MuZero on the pixel-based cartpole-balance environment, obtained by running LightZero/zoo/dmc2gym/config/dmc2gym_pixels_smz_config.py:

From the results, it can be seen that the implementation's effectiveness is basically validated. For baseline results on other DMC environments with pixel-based input, it is expected to be similar to those with state-based input. You may try further testing locally. Thank you for your support!