Metaworld
Metaworld copied to clipboard
Collections of robotics environments geared towards benchmarking multi-task and meta reinforcement learning
Meta-World was designed to be both a Meta-RL and a Multi-Task RL benchmark. One of the awkward consequences of that is that the way goal conditioning is handled is very...
The code used to create `metaworld.envs.ALL_V2_ENVIRONMENTS_GOAL_OBSERVABLE` and `metaworld.envs.ALL_V2_ENVIRONMENTS_GOAL_HIDDEN` is highly non-obvious and among other things involves running a regex on the environment name. It should be possible to implement this...
This rarely matters, but `button-push` family of tasks also reward pulling on the button. This makes it possible to get very high reward without ever succeeding at the task, which...
In discussions with a PhD college, they were explaining that life long learning in RL doesn't have a set of environments, API, wrappers, etc for researchers to use. Therefore, I...
- Type hinted all envs, utils and policies - Added static type checking to pre-commit with MyPy - Added some type conversions to some env internals to ensure static type...
This PR re-organizes the Meta-World repository by: - removing the version 1 environments (we can keep the reward functions by adding a per environment flag about which reward function to...
The reward function currently assigns a higher reward when the gripper is merely against the handle compared to when it is actually hooking the handle. Additionally, the reward when the...