Skip to content

Latest commit

 

History

History
13 lines (8 loc) · 1.36 KB

README.md

File metadata and controls

13 lines (8 loc) · 1.36 KB
Raw

This repository provides software to construct skills using pre-trained Hierarchical Actor-Critic (HAC) agents. As an example, executing the command python3 run_random_agent.py --show --subgoals will implement an ant agent in a four rooms enivronment that randomly executes a set of three skills, which include

  1. Move to northern door of room
  2. Move to center of room
  3. Move to southern door of room.

Please note that in order to load the pre-trained agents, the file paths listed in the "checkpoint" file within the "models" folder need to be replaced with the current file paths. That is, the paths "/Users/andrewlevy/.../models/HAC.ckpt-99" should be replaced with "/YourFilePath/models/HAC.ckpt-99".

To visualize the skills without the HAC subgoals, remove the --subgoals option. To train agent without the visualization, remove the --show option. The following video shows a preview of these skills.

Skills can be constructed/modified using the execute_skill.py, agent.py, and environment.py files.

Please note that in order to run this repository, you must have (i) a MuJoCo license, (ii) the required MuJoCo software libraries, (iii) the MuJoCo Python wrapper from OpenAI, and (iv) TensorFlow.