1. Car Racing Competition

2. Goal  Learn or design a controller for TORCS that races as fast as possible alone or in the presence of others drivers  “Spiritual successor” of CEC 2007 Competition

3. Car Racing Competition meets TORCS  More representative of real game AI  Better interaction with human players  Many good programmed controllers available  Challenges How to make it easy accessible? Not designed for Machine Learning! More similar to a real-world problem

4. The Open Racing Car Simulator

5.  TORCS is a state of the art open source simulator written in C++  Main features Sophisticated dynamics Provided with several cars, tracks, and controllers Active community of users and developers Easy to develop your own controller  OS Support Linux: binaries and building from sources suppo Windows: binaries and “limited” bulding from sources support OSX: legacy binaries and no building from sources support 

6. Competition API  To make TORCS more easy to use we developed an API based on socket (UDP)  Values of sensors and effectors are sent through UDP  3 components Torcs Patch Server Bot (C++) Client API (C++ and Java) Server BOT TORCS Client Controlller Client Controlller Patch UDP

7. Sensors Track SensorsOpponent Sensors  Rangefinders to sense Opponents Edges of the track  Speed  Position on track  Rotation speed of wheels  RPM  Angle with track  Distance raced  Fuel and damage ...

8. Effectors  Basically 4 effectors Steering wheel [-1,+1] Gas pedal [0, +1] Brake pedal [0,+1] Gearbox {-1,0,1,2,3,4,5,6}

9. Rule  Hand coding controller were allowed  Unknown Track  Distance raced by each controller within 10000 tics(200 sec)

10. The Results

11. Submissions  5 entries have been submitted to the competition: Matt Simmerson Leonard Kinnaird-Heether – Wayne State University Chin Hiong Tan – National University of Singapore Diego Perez - University Carlos III, Madrid Simon Lucas – University of Essex  3 more controllers have been considered Daniele’s heuristic C++ controller Julian’s heuristic Java controller Luigi Cardamone - Politecnico di Milano

12. Scoring setup  A server with 2 Quad-core Xeon 2.66 GHz, 8GB RAM, running Fedora Core 6

13. Scoring process  Scoring is a two stage process involving three tracks (unknown to the competitors): RUUDSKOGEN STREET1 D-SPEEDWAY  In the first stage only a controller at once is tested and performance is defined as the distance covered within 10000 game tics  In the second stage all the controllers (except the three not in the competition) race at the same time and performance is defined simply as the arrival order after 3 laps

14. First Stage: RUUDSKOGEN

15. First Stage: STREET1

16. First Stage: D-SPEEDWAY

17. First Stage: overall results  F1 point system is used to compute the final scores of the controllers on the 3 tracks

18. Second Stage: Competition League  Also in this stage, F1 point system is used to compute the final score

19. Second Stage: Full League

20. Summary of results  3 learned controllers are significantly better than programmed controllers  Some possible overfitting to training tracks?  Poor performance in avoiding and overtaking opponents  Some problems with recovery behavior  Still more work to reach the performance of controllers provided with TORCS but initial results are promising!

21. TORCS  Official Site http://torcs.sourceforge.net/  Source Code Download http://torcs.sourceforge.net/index.php?name= Sections&op=viewarticle&artid=3  Robot Tutorial http://www.berniw.org/