Take MsPacman-v0 for example. I know we can use env.action_space.sample() to move agent in random direction. But how can I move the agent in the directions as I desired?
you will need to figure out the action id -> joystick direction mapping (consult https://github.com/openai/gym/blob/master/gym/envs/atari/atari_env.py#L163)
and pass the desired action id to env.step(action_id)