I visualized the state coverage of the provided data in the 2D maze environments. For the small UMaze environment the downloadable dataset seems to have a skewed data distribution where the agent never fully explores one of the sides:
The right shows a scatter plot of all positions in the downloaded dataset. Was there a bug in data generation or is this intended? (the data that is used to generate the GIF on the website does not seem to have this issue)
The other two maze environments seem to have full coverage of the maze:
I slightly modified the visualized_dataset.py script for these plots:
import argparse
import d4rl
import gym
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--env_name', type=str, default='maze2d-umaze-v0')
args = parser.parse_args()
env = gym.make(args.env_name)
dataset = env.get_dataset()
if 'infos/qpos' not in dataset:
raise ValueError('Only MuJoCo-based environments can be visualized')
qpos = dataset['infos/qpos']
qvel = dataset['infos/qvel']
rewards = dataset['rewards']
actions = dataset['actions']
import matplotlib.pyplot as plt
plt.scatter(qpos[:, 0], qpos[:, 1])
plt.axis('equal')
plt.show()
I visualized the state coverage of the provided data in the 2D maze environments. For the small UMaze environment the downloadable dataset seems to have a skewed data distribution where the agent never fully explores one of the sides: The right shows a scatter plot of all positions in the downloaded dataset. Was there a bug in data generation or is this intended? (the data that is used to generate the GIF on the website does not seem to have this issue)
The other two maze environments seem to have full coverage of the maze:
I slightly modified the
visualized_dataset.py
script for these plots: