Large

Description

The data is collected from the PointMaze_Large-v3 environment. The agent uses a PD controller to follow a path of waypoints generated with QIteration until it reaches the goal. The task is continuing which means that when the agent reaches the goal the environment generates a new random goal without resetting the location of the agent. The reward function is sparse, only returning a value of 1 if the goal is reached, otherwise 0. To add variance to the collected paths random noise is added to the actions taken by the agent.

Dataset Specs

Total Timesteps	1000000
Total Episodes	3071
Flatten Observations	True
Flatten Actions	False
Algorithm	QIteration
Author	Rodrigo Perez-Vicente
Email	rperezvicente@farama.org
Code Permalink	https://github.com/rodrigodelazcano/d4rl-minari-dataset-generation
download	minari.download_dataset("pointmaze-large-v0")

Environment Specs

ID	PointMaze_Large-v3
Action Space	Box(-1.0, 1.0, (2,), float32)
Observation Space	Dict('achieved_goal': Box(-inf, inf, (2,), float64), 'desired_goal': Box(-inf, inf, (2,), float64), 'observation': Box(-inf, inf, (4,), float64))