René's URL Explorer Experiment

{"@context":"https://schema.org","@type":"DiscussionForumPosting","headline":"Meta-World","articleBody":"![image](https://user-images.githubusercontent.com/1320252/126459115-df7fb60c-f8c7-4e99-8eca-857fe23d8696.png)\r\n\r\nMeta-World is an open-source simulated benchmark for meta-reinforcement learning and multi-task learning consisting of 50 distinct robotic manipulation tasks.\r\n- https://github.com/rlworkgroup/metaworld\r\n- https://arxiv.org/pdf/1910.10897.pdf\r\n\r\n## The actions in this space \r\n\r\nrange between −1 and 1. \r\n\r\nFor all tasks, the robot must either manipulate \r\n- one object with a variable goal position, or \r\n- manipulate two objects with a fixed goal position.\r\n\r\n## The observation space is represented as \r\n\r\n- a 6-tuple of the 3D Cartesian positions of the end-effector, \r\n- a normalized measurement of how open the gripper is, \r\n- object\r\n  - the 3D position of the first object, \r\n  - the quaternion of the first object, \r\n  - the 3D position of the second object, \r\n  - the quaternion of the second object, \r\n- all of the previous measurements in the environment, and \r\n  - finally the 3D position of the goal. \r\n\r\nIf there is no second object or the goal is not meant to be included in the observation, then the quantities corresponding to them are zeroed out. The observation space is always 39 dimensional.","author":{"url":"https://github.com/NirViaje","@type":"Person","name":"NirViaje"},"datePublished":"2021-07-21T09:12:01.000Z","interactionStatistic":{"@type":"InteractionCounter","interactionType":"https://schema.org/CommentAction","userInteractionCount":0},"url":"https://github.com/5/DeepLearningRobotics/issues/5"}