Construction environments is a set of reinforcement learning environments with the task of building complex objects from simpler ones. Construction environments is a Unity project inplemented using the Unity ml-agents tookit.
A reinforcement learning agent needs to learn how to build a toy bridge. Goal: Link the start (blue) and target (red) cubes using the given basic blocks - a support, a square plank, 2 long planks. Actions: An agent can choose 4 directions where to move and which building block to place. Observations: Coordinates of blue and red cubes, an array with nearby objects.
Examples of training using Unity ml-agents toolkit built-in PPO algorithm.
| After 10 000 steps | After 1.5 million steps | After 3 million steps |
|---|---|---|
 |
 |
 |