Simulated High-Speed Vehicle Control

We have applied Dynamic Threshold Learning to high-speed vehicle control behaviors using the open source TORCS (The Open Race Car Simulator) simulator.  DTL is able to quickly learn dynamic stability control (DSC) behaviors that prevent understeer and oversteer on a wide variety of terrain types – from fast speedways with sweeping curves, to twisting mountain roads, to tight dirt tracks.  Unlike conventional DSC systems on automobiles, which only react once wheel slip is detected, DTL learns the appropriate steering and acceleration  commands to prevent understeer and oversteer from occurring in the first place.

This video shows DTL learning DSC behaviors on a twisting alpine road:

In less than a minute, DTL learns the parameters needed to avoid both understeer and oversteer. Additional details are available in our paper submitted to ICRA 2012.