Mujoco-sim

Learning environments for robotic manipulation using MuJoCo with the dm_control package. (this repo also serves as project for learning to use both myself, so not everything is implemented in the best/ most canonical way).

Tasks

Robot Tasks

All these tasks have both dense and sparse rewards, and both visual and state observations.

Reach The agent has to reach a target location in the Euclidean space by providing delta steps for the position (xyz).

Planar Push The agent controls delta xy coordinates of a cylindrical EEF and has to push a configurable number of objects to a target location (indicated by a white disc).

repo layout

/mujoco_sim
    /entities           # all phyiscal objects in the environments, using the composer.Entity abstraction
        /arenas         # roots of the entity tree, contain the 'robot setup'
        /robots         # actual robots
        /eef            # grippers etc.
        /props          # non-actuated elements
    /environments
        /tasks          # implements the actual learning tasks
        dmc2gym.py      # converts the DMC environment interface to a gym interface for interacting with most RL frameworks
    /mjcf               # contains the mjcf xml files for all the entities

installation

git clone
git submodules update --init
conda env create -f environment.yaml
pip install -e ur_ikfast/
pip install -e airo-core/

for learning:

pip install -e .[sb3]

Dependencies and information sources

Dependencies

mujoco_menagerie: library of MJCF models
ur_ikfast: python ikfast wrapper for UR robot class