search

fpadula / visualcollisionarm

End-to-End Visual Obstacle Avoidance for a Robotic Manipulator usingDeep Reinforcement Learning
8 stars 3 forks source link

Where does the training occurs? #3

Open fpadula opened 2 months ago

fpadula commented 2 months ago

(From here)

Hi @Hokite, please open an issue ticket next time. The learning part is not performed on Unity itself but rather on Python. I use a framework called ml-agents that provides an abstraction layer to control an environment running in Unity/C# from within Python. Very similar to what a gym environment does (if you are familiar with that).

Hokite commented 2 months ago

Hi @fpadula, I enjoyed your reinforcement learning research. My question is, I would like to ask about the specific status, behavior, and rewards in your learning. can you tell me the reinforcement learning C# script you used to create unity gym?

fpadula commented 2 months ago

Hi @fpadula, I enjoyed your reinforcement learning research. My question is, I would like to ask about the specific status, behavior, and rewards in your learning. can you tell me the reinforcement learning C# script you used to create unity gym?

Hi @Hokite, you can find information about the modeling, rewards, states, etc, in my Master's thesis here: https://www.researchgate.net/publication/358468389_End-to-End_Visual_Obstacle_Avoidance_for_a_Robotic_Manipulator_using_Deep_Reinforcement_Learning

As for the C# script, in order to create the Gym interface you have to implement a few scripts as far as I remember, but most of the agent code is in https://github.com/fpadula/visualcollisionarm/blob/master/Unity_project/Assets/Scripts/EEAgent.cs

I recommend taking a look at https://github.com/Unity-Technologies/ml-agents to understand better how everything is structured.