Training a virtual bipedal robot to walk in OpenAI Gym using Evolution Strategies (ES).
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
README.md
agent.py
agent_play.py
agent_train.py
weights.pkl

README.md

bipedalrobot

Training a virtual bipedal robot to walk in OpenAI Gym using Evolution Strategies (ES).

Credit goes to alirezamika for writing most of the code. I added documentation and optimized hyperparameters.

Feel free to tweak the hyperparameters in agent.py and see if you get a better reward. The currently trained model gets a total reward of ±190.

Example video of one episode:

IMAGE ALT TEXT HERE

Dependencies

To use this model you have to install OpenAI Gym and the Evostra repository of alirezamika:

pip install evostra
pip install gym

Training

Type in command line:

python agent_train.py

The program trains standard for 200 iterations. If you want you can change the amount of iterations in agent_train.py by changing the 'n' variable.

Testing / Playing

Type in command line:

python agent_play.py

The program plays one episode. If you want to change the number of episodes, change the 'n' variable in agent_play.py.

Study material

This repository was created as a midterm assignment for School of AI's "Move 37" course.

For more information about the "Move 37" course check out: https://www.theschool.ai/courses/move-37-course

For more information about Evolution Strategies (ES) in general: https://blog.openai.com/evolution-strategies