This directory contains examples for running MPC algorithms on gym-like environments.

1. Clone and checkout the below repos by following setup instructions in individual repos:

   • mjrl -- has detailed setup instructions here
   • trajopt -- this repository containing the MPC algorithms

2. Run a toy experiment (point mass) using:

   python job_script.py --output point_mass_job --config configs/point_mass_config.txt
   

   This optimization should take only a second on so. After the optimization is complete, there will be a prompt to visualize the optimized trajectory. This can be turned off by changing the debug field in the config file.

3. Make a 7DOF sawyer arm reach various goals with the end effector (either fixed goal in episodic setting, or changing goals in continual setting)

   python job_script.py --output reacher_7dof_job --config configs/reacher_7dof_config.txt
   python job_script.py --output continual_reacher_7dof_job --config configs/continual_reacher_7dof_config.txt
   

4. In-hand manipulation of a pen with a Shadow Hand (Adroit). This requires installation of the mj_envs repository. This package uses git submodules, additional care must be taken to exactly follow setup instructions. NOTE: Uncomment the import mj_envs line in the job_script.py file.

   python job_script.py --output pen_job --config configs/pen_config.txt
   

5. To visualize previous runs (stored as pickle files)

   python visualize_trajectories.py --file <path/to/file.pickle> --repeat <#times to repeat>