d4rl-atari
d4rl-atari copied to clipboard
How are actions in the dataset defined?
To be more specific, in the ms-pacman environment, I use env = gym.make('ms-pacman-medium-v0')
to create the env, and I run
observation = env.reset() # observation.shape == (84, 84)
observation, reward, terminal, info, _ = env.step(2)
I found that the action definition here is the same as https://gymnasium.farama.org/environments/atari/ms_pacman/ , action 2 controls the character to move to the right.
But in data = env.get_dataset()
, I found that there seems to be some difference.
data['actions'][:10]
array([2, 2, 2, 2, 6, 2, 2, 7, 7, 7], dtype=int32)
I saved the first few frames of observation:
Why is the character moving upwards? Am I missing something? @takuseno @takuyamagata