Skip to content

Latest commit

 

History

History
43 lines (31 loc) · 2.66 KB

readme.md

File metadata and controls

43 lines (31 loc) · 2.66 KB
Raw

GTAV RewardHook

GTAV RewardHook is a plugin for ScriptHookDotNet that turns GTA V into a reinforcement learning environment.

The idea is to accumulate facts relevant to safe driving (such as average road alignment, collisions etc.) in each training episode. These stats can then be translated to a single reward using some configurable reward scheme, and can be stitched together with screenshots taken around the same time to create datasets for reinforcement learning agents.

Disclaimer

The publisher of GTA V allows non-commercial use of footage from the game. Please do not use any datasets generated using RewardHook for any commercial projects. This work is also not affiliated with any organization that I am a part of and is just a personal project :)

Events

RewardHook is currently able to record the following accurately in each episode:

  • Average speed
  • Average road alignment (Measured as the cosine of the angle between road and vehicle)
  • Collisions with vehicles
  • Collisions with pedestrians
  • Collisions with other objects / environment
  • Driving against traffic
  • Driving on sidewalks

Getting Started

Requirements

  1. GTA V (with ScriptHook and ScriptHookDotNet installed)
  2. ScriptHookDotNet (for compilation, from nuget, VS should get this automatically)
  3. Nancy (for the RESTful API, from nuget, VS should get this automatically)

Installation

  1. Download / clone the repository.
  2. Open the project, and build using Visual Studio (preferably using VS2017 or later).
  3. Drag the compiled dll to scripts in your GTAV root.
  4. Start the game. You should now see occasional notifications.

Pending Tasks

RewardHook is still a work in progress. While a lot of events can be extracted using the known list of native methods (from GTA's RAGE engine), there are still several events that we cannot reliably detect (like running red lights, or driving offroad, or even lane alignment), unless the corresponding native addresses are known (or exist).

For some of these events, alternative tracking mechanisms do exist. For example, it is currently possible to detect if an NPC is waiting at a red light, which can help us create datasets that penalize running red lights only when another NPC is waiting at a red light and has been travelling on the same road.

Fortunately, the list of natives is constantly growing as the modding community discovers more natives (I use this list as a reference).

Besides these limitations, any pending tasks are usually in TODO.md.

Contributing

Contributions are welcome. Feel free to send a pull request.