/TorcsTool

torcs client for deep learning or reinforcement learning

Primary LanguageC++MIT LicenseMIT

Welcome to TorcsTool

TorcsTool is a torcs client for reinforcement learning like gym_torcs. The client developed by udp has some delay when I grab images from Torcs, so TorcsTool develped by shared memory to boost image grab. I modify scr_server in drivers.

This repo is for torcs-1.3.7.

Getting Started

  1. Download torcs-1.3.7.

  2. Create git repo, and apply patch.

    git apply --ignore-space-change --ignore-whitespace  torcs.patch

  3. Compile and install torcs

  4. Compile and install scr-server in your torcs folder.

  5. Compile shared library.

    cd path/to/TorcsTool && make

  6. Run tool.py and torcs

Api

get29Data

return 29 track data collected from scr_server.

  1. angle: angle between track and car.
  2. track: distance from car to track edge in 19 direction. You can modify direction in src_server.cpp about line 246 in my case.
  3. speed(x|y|z): speed in x,y,z axis.
  4. wheelSpinVel: rad speed of wheel
  5. rpm: rpm

image

return images of torcs. Default grab speed is 10 images per second. You can change grab speed by modifying count in src/libs/raceengineclient/raceengine.cpp about 767 line in my case.

accel/steer/brake/clutch/gear

control car, just set the value.

restart

restart game.

is_stuck/is_hit_wall/is_finish

check car status. You can modity ReOneStep function in src/libs/raceengineclient/raceengine.cpp to change the status trigger.

Dependency

  1. torcs-1.3.7
  2. python3
  3. see requirement.txt

License

TorcsTool is released under the MIT License.