Large-Play

Description

The data is collected from the AntMaze_Large-v4 environment. At the beginning of each episode random locations for the goal and agent’s reset are selected. The success rate of all the trajectories is more than 80%, failed trajectories occur because the Ant flips and can’t stand up again. Also note that when the Ant reaches the goal the episode doesn’t terminate or generate a new target leading to a reward accumulation. The Ant reaches the goals by following a set of waypoints using a goal-reaching policy trained using SAC.

Dataset Specs

Total Steps	1000000
Total Episodes	1000
Dataset Observation Space	Dict('achieved_goal': Box(-inf, inf, (2,), float64), 'desired_goal': Box(-inf, inf, (2,), float64), 'observation': Box(-inf, inf, (27,), float64))
Dataset Action Space	Box(-1.0, 1.0, (8,), float32)
Algorithm	QIteration+SAC
Author	Alex Davey
Email	alexdavey0@gmail.com
Code Permalink	https://github.com/rodrigodelazcano/d4rl-minari-dataset-generation
Minari Version	0.4.3 (supported)
Download	minari download antmaze-large-play-v1

Environment Specs

Note

The following table rows correspond to (in addition to the action and observation space) the Gymnasium environment specifications used to generate the dataset. To read more about what each parameter means you can have a look at the Gymnasium documentation https://gymnasium.farama.org/api/registry/#gymnasium.envs.registration.EnvSpec

This environment can be recovered from the Minari dataset as follows:

import minari

dataset = minari.load_dataset('antmaze-large-play-v1')
env  = dataset.recover_environment()

ID	AntMaze_Large-v4
Observation Space	Dict('achieved_goal': Box(-inf, inf, (2,), float64), 'desired_goal': Box(-inf, inf, (2,), float64), 'observation': Box(-inf, inf, (27,), float64))
Action Space	Box(-1.0, 1.0, (8,), float32)
entry_point	gymnasium_robotics.envs.maze.ant_maze_v4:AntMazeEnv
max_episode_steps	1000
reward_threshold	None
nondeterministic	False
order_enforce	True
autoreset	False
disable_env_checker	False
kwargs	{'maze_map': [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1], [1, 0, 1, 1, 0, 1, 0, 1, 0, 1, 0, 1], [1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1], [1, 0, 1, 1, 1, 1, 0, 1, 1, 1, 0, 1], [1, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 1], [1, 1, 0, 1, 0, 1, 0, 1, 0, 1, 1, 1], [1, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]], 'reward_type': 'sparse', 'continuing_task': True, 'reset_target': False}
additional_wrappers	()
vector_entry_point	None

Evaluation Environment Specs

Note

This environment can be recovered from the Minari dataset as follows:

import minari

dataset = minari.load_dataset('antmaze-large-play-v1')
eval_env  = dataset.recover_environment(eval_env=True)

ID	AntMaze_Large-v4
Observation Space	Dict('achieved_goal': Box(-inf, inf, (2,), float64), 'desired_goal': Box(-inf, inf, (2,), float64), 'observation': Box(-inf, inf, (27,), float64))
Action Space	Box(-1.0, 1.0, (8,), float32)
entry_point	gymnasium_robotics.envs.maze.ant_maze_v4:AntMazeEnv
max_episode_steps	1000
reward_threshold	None
nondeterministic	False
order_enforce	True
autoreset	False
disable_env_checker	False
kwargs	{'maze_map': [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 'r', 0, 0, 0, 1, 0, 0, 0, 0, 0, 1], [1, 0, 1, 1, 0, 1, 0, 1, 0, 1, 0, 1], [1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1], [1, 0, 1, 1, 1, 1, 0, 1, 1, 1, 0, 1], [1, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 1], [1, 1, 0, 1, 0, 1, 0, 1, 0, 1, 1, 1], [1, 0, 0, 1, 0, 0, 0, 1, 0, 'g', 0, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]], 'reward_type': 'sparse', 'continuing_task': True, 'reset_target': False}
additional_wrappers	()
vector_entry_point	None