Skip to content

KisangPark/Ant_control

Repository files navigation

Mujoco Ant Reinforcement learning control

The necessity of whole-body manipulation is gaining attention due to the increasing usage of quadruped robots in various industrial situations. To overcome this complex problem, which can possibly contain lots of contacts, new and effective methods must be tried. This work presents a new perspective of dealing with the whole-body manipulation problem of quadruped Ant robot, applying end-to-end network and multi-agent reinforcement learning. By using these methods, quadruped Ant robot achieved an improvement in total execution time and stability. Furthermore, it was possible to obtain a new and unexpected behavior of Ant robot in multi-agent learning.

Multi-agent controller structure

image

DDPG, centralized critic

DDPG algorithm applied
Global reward & single critic network
image

Results

image

Rendering Video

--> Unexpected new form of behavior discovered

MACC_success.mp4

Releases

No releases published

Packages

No packages published