Features for 0.9 release

[edbeeching]

My ideas for the 0.9 release are the following:

• Deprecation of using gdrl to launch training, I think it is better to have users refer to and use the examples as the sole way to use the lib.
• Creation of a "pretrained behavior library" that will contain some basic 2D and 3D pretrained models for simple tasks such as point-to-point navigation, patrolling, exploration and enemy avoidance. The creation of this is quite an endeavor, so I plan to start with a MVP with just the point-to-point task. The objective of this feature is to allow users to add behaviours to their NPCs without the need to train/ install the python part of the plugin.
• Creation of a dedicated Hugging Face org to store example environments and agent behaviors: https://huggingface.co/godot-rl-agents
• Add more tests.
• Add an example with an agent interacting in a discrete environment, I am not even sure if this would sync correctly, needs testing.

[Ivan-267]

One suggestion until/for 1.0 (not necessarily 0.9) is to extend the API with terminated/truncated and set the default AIController timeout restart to truncated, as it might have an effect on training.

[Ivan-267]

Add an example with an agent interacting in a discrete environment, I am not even sure if this would sync correctly, needs testing.

We now have an example of a discrete environment at https://github.com/edbeeching/godot_rl_agents_examples/tree/main/examples/CrossTheRoad.

We can consider other examples in the future. Envs that require manual/async stepping are not directly supported and may be more challenging without further changes.

[GeorgeS2019]

With the Godot 4.3, we now have a complete new rewritten skeleton3D framework.

It would be great to consider Mujoco environment for 0.9 or future release of Godot RL Agents