- Install
requirements.txt cd environments/customenv/mujococustom && git clone https://github.com/google-deepmind/mujoco_menagerie.git assets
You might need to use an LD_PRELOAD trick if you're using conda. Here: https://stackoverflow.com/questions/71010343/cannot-load-swrast-and-iris-drivers-in-fedora-35
We provide a training environment that loads ten arms at once. You can use the files in config/multienv and config/env
to control which arms are loaded, what amount of domain randomization is applied, as well as how many parallel simulators per arm are loaded.
Code a policy along with associated training and deployment wrappers.
You can use any number of the arms mentioned in config/env/all-with-dr.yaml.
You may not use the wx250s as part of training.
Your goal is to obtain as high performance as possible on the wx250s.
The task your policy must solve is to reach a target end effector position (x=0.5, y=0, z=0.3).
You are allowed to use the wx250s for evaluation purposes, as defined in config/env/allandwidow-with-dr.yaml.
Indeed, you will notice that in config/multienv/multienv.yaml, all arms (except wx250s) are defined as training environments
and all arms (including wx250s) are defined as evaluation environments.
We will host a friendly competition.
If our resources permit, a physical wx250s will be on premises,
and we will deploy your policy on it.
During this deployment, the API from your policy's POV will be the same as during training: you will receive a wx250s
observation with the same shape as is used in this simulator.