- This repository provides the tools to train a Jaco2 arm with dm_control. It depends on our version of dm_control which has network hooks for collecting real robot state and commanding actions.
- The agent implementations are taken from the Official TD3 repository.
python main.py --domain jaco --task relative_position_reacher_7DOF --policy TD3 --seed 100 --device 'cuda:0' --exp_name 'jaco_relative_reacher'
python main.py --domain jaco --task relative_position_reacher_7DOF --policy TD3 --load_model results/jaco_relative_reacher_00/ --plot_action_movie
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.
python main.py --domain reacher --task easy --policy TD3 --seed 100 --device 'cuda:0' --exp_name 'reacher_easy'
python main.py --domain reacher --task easy --policy TD3 --load_model 'results/reacher_easy_00' --eval --plot_movie
Debugging doc with details on robot positions and angles.
This work was built off of a collection of repos developed by @sahandrez