Skip to content
This repository was archived by the owner on Dec 18, 2024. It is now read-only.

README.md

Latest commit

 

History

History
40 lines (30 loc) · 1.72 KB

README.md

File metadata and controls

40 lines (30 loc) · 1.72 KB

Reproduce ContactInsert

Installation

We use MuJoCo 2.0.0 for physics simulation. After following the general SAC Utilities installation steps, do the following:

Common to both Linux and MacOS

  • Download MuJoCo 2.0.0 in ~/.mujoco. Make sure the dicrectory structure is ~/.mujoco/mujoco200_linux/{bin,doc,include,model,sample}.
  • cp mjkey.txt ~

Extra steps for MacOS

  • brew install gcc
  • export CC=/usr/local/bin/gcc-11 (check the GCC version actually installed by Homebrew and modify accordingly)
  • Go to System Preferences -> Security and Privacy -> Developer Tools. Give your terminal app permission to run software locally does not meet the system's security policy.
  • Navigate to the mujoco200_linux directory, right click all .dylib files and click Open -> Allow.

Continued: Common to both Linux and MacOS

  • cd install_contact_insert && ./install.sh ../../robosuite
  • conda activate sac_utils

Training

ContactInsert training curve

(base) $ conda activate sac_utils
(sac_utils) $ ./run_script.sh training_variables_contact_insert.sh

Policy Evaluation

Download the checkpoint from links below, and unzip it at CKPT_DIR.

Model Link CKPT_NAME Experiment ID
ours Google Drive policy_checkpoint_0005200000 exp_137 
(sac_utils) $ python tfagents_system/run.py --params tfagents_system/contact_insert_params.py --exp_dir $CKPT_DIR \
--config config/contact_insert.json --num_episodes 5 --checkpoint $CKPT_NAME