Adding save and load support

[rimvydasrub]

Currently, its not straightforward how the save and load functionality works for the environment. This functionality is valuable for:

• replication of the results as just setting random seed (random.seed(###)) is not sufficient to fix randomness in the system
• go generate environments and do feature prepossessing offline so that training and inference can be done in a faster manner.

[peterdavidfagan]

Can you provide further details, what are you asking for in the API? What are you looking to save/load?

Reproducibility and random seeds is something I can look into at some point.

[rimvydasrub]

If you run the following program

import random

from PIL import Image
import numpy as np
import torch
from mujoco_robot_environments.tasks.rearrangement import RearrangementEnv

# setting random seeds
random.seed(0)

env = RearrangementEnv(viewer=True)
env.reset()

obs = env._compute_observation()['overhead_camera/rgb']
obs_img = Image.fromarray(obs)
obs_img.save('obs.png')

env.close()

you get two different results

After inspection, this is because np.random.seed is not set explicitly. I can now replicate environments with the same seed.

Generally, for the API functionality, is something like this possible?

from mujoco_robot_environments.tasks.rearrangement import RearrangementEnv

env = RearrangementEnv(viewer=True)
env.reset()

# saving the environment
env.save("save_file_name")

# loading the environment
env = RearrangementEnv.load("save_file_name")