chernyakkonstantin / godot_gym_api Goto Github PK

Consider default repeat_action = 4.

My robot uses Lidar sensor with maximum data volume of 1080 points per repeat. In total I have 4320 points. The rest data is small enough to not consider them.

One environment step takes 0.14s, out of which ~0.06s is spent on filling protobuf message (Lidar data only) and 0.05s is spent on serialization of entire protobuf message.

It is required to test, whether the following algorithm is faster:

1. Make message to contain data per one repeat.
2. Fill, serialize, and send data in separate threads at each repeat.
3. In python, accumulate data over repeats to form data per step.

The expected improvement upper bound is that repeat_action times faster than current implementation.

Need to implement multi-agent support (e.g. two agents play hide-and-seek in the same arena).

User should be capable to set time in a game to be faster than real time.

It is required to obtain more samples per second.

Need for CLI commands to install Godot add-on in specified directory if a user has installed Python package.

Need for solution that works on Windows, MacOS, Linux.

User should be capable to select between TCP / Shared Data (e.g. to run a Godot application on a remote machine).

Need to implement running several same environments in parallel.

First, prove hypothesis that running multiple Godot applications consume more resources than running single application with multiple scenes.

This will allow to gather more samples per second.

Need to create plugins to easier implementing different tasks.

E.g. set of sensors to use as observations for a self-driving car.