TRC: Trust Region Conditional Value at Risk for Safe Reinforcement Learning

This is an official GitHub Repository for paper (link):

• Dohyeong Kim and Songhwai Oh, “TRC: Trust region conditional value at risk for safe reinforcement learning,” IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 2621–2628, Apr. 2022.

• This paper proposes a CVaR-constrained safe RL method (called TRC).

License

Distributed under the MIT License. See LICENSE for more information.

requirement

• python 3.7 or greater
• gym
• mujoco-py (https://github.com/openai/mujoco-py)
• safety-gym (https://github.com/openai/safety-gym)
• stable-baselines3
• tensorflow-gpu==1.13.1
• tensorflow-probability==0.6.0
• torch==1.10.0
• requests
• wandb

How to use

tf1

• training:

  • cd tf1
    bash train_{env_name}.sh

• test:

  • cd tf1
    bash test.sh

torch

• training:

  • cd torch
    bash train_{env_name}.sh

• test:

  • cd torch
    bash test.sh

Supported environments

• Safety-Gym: Safexp-PointGoal1-v0, Safexp-PointGoal1-v0, Doggo-v0 (which is a hierarchical version of Safexp-DoggoGoal1-v0)

• MuJoCo: Jackal-v0