Robust Grasping

**Reimplementation of "Robust Adversarial RL" Paper (https://arxiv.org/abs/1703.02702)

Requirements

• mujoco_py==0.5.0
• mjpro131
• Fork of OpenAI gym [https://github.com/lerrel/gym-adv]
• Tensorflow>=1.5
• Latest OpenAI baselines

Usage

• [MPI multiprocessing] (mpi installed)

  • Enable 8 workers: mpirun -np 8 python main_mujoco_robust.py

• [PPO & Robust Agent]

  • PPO.py
  • PPO_RARL.py
  • Policy network: MlpPolicy.py

• [Training]

  • Train robust agent: python main_mujoco_robust.py
  • Train ppo agent: python main_mujoco_ppo.py
  • Eval agent: python eval.py

• [Visualization]

  • Plot with variation python visualize.py

usage: run_mujoco_ppo.py [-h] [--env ENV] [--seed SEED] [--sr SR]

optional arguments:
  -h, --help   show this help message and exit
  --env ENV    environment ID (default: HopperAdv-v1)
  --seed SEED  RNG seed (default: 10)
  --sr SR      success reward (default: 3800.0)

To-do list

• Add rendering code for adversarial force in gym-adv
• Add new robotic manipulation environment