RL on Jaco Arm in DeepMind Control Suite

• This repository provides the tools to train a Jaco2 arm with dm_control. It depends on our version of dm_control.
• The agent implementations are taken from the Official TD3 repository.

Example training for Jaco2 7DOF robot with position controllers:

python main.py --domain jaco --task relative_position_reacher_7DOF --policy TD3 --seed 100 --device 'cuda:0' --exp_name 'jaco_relative_reacher'

Example plot and eval latest trained model

python main.py --domain jaco --task relative_position_reacher_7DOF --policy TD3 --load_model results/jaco_relative_reacher_00/ --plot_action_movie

Example plot and eval latest trained model with topview camera

python main.py --domain jaco --task relative_position_reacher_7DOF --policy TD3 --load_model results/jaco_relative_reacher_00/ --plot_action_movie --camera_view 'topview'

Example gif from trained reacher where first 7 color bars are relative actions of joints 1-7 and the final red bar is reward.

Example training for reacher:

python main.py --domain reacher --task easy --policy TD3 --seed 100 --device 'cuda:0' --exp_name 'reacher_easy'

Example plot and eval previous experiment

python main.py --domain reacher --task easy --policy TD3 --load_model 'results/reacher_easy_00'  --eval --plot_movie

General tips:

Debugging doc with details on robot positions and angles.

Thanks

This work was built off of a collection of repos developed by @sahandrez