About Transfer Learning

[abdulkadrtr]

Quadruped Robot Navigation with ML-Agents

I am working on a quadruped robot navigation project using Unity and ML-Agents. The main objective is to teach the robot how to walk, followed by teaching it to navigate in obstacle-rich environments. I plan to achieve this in stages.

Initial Setup

• The observation space will have 20 elements.
• The action space will consist of 6 elements.

At the beginning, when teaching the robot to walk, I will use a 14-element observation space and set the remaining neurons to a default value of -1. Once the model is trained, I plan to use transfer learning by initializing the model with --initialize-from. In the new scenario, instead of setting the remaining 6 elements to -1, I will provide the actual observation values.

Is this approach suitable?

Is this a good approach to solve my problem, or is there an alternative way to handle this situation more effectively?