Dataset and Benchmark
Seven maps with various levels of difficulty are designed for the drifting task, for which we refer the tracks of a racing game named PopKart. The difficulty of driving increases from (a) to (g). Maps (a-f) are used for our training and (g) is for evaluation. They are generated by RoadRunner, a road and environment creation software for automotive simulation. All maps and the code scripts for the benchmark tests are availiable. And here we provide the benmchmarks from our experiments on different performance metrics.
- Performance Metrics
C.T.E. and H.A.E.: cross track error and heading angle error.
MAX-VEL and AVG-VEL: the maximum and average velocity of a driving test.
SLIP: the maximum slip angle during a driving test.
L.T.: the time to reach the destinations (the lap time).
SMOS: the smoothness of driving, calculated by the rolling standard deviation of steering angles.
- Benchmark
In the paper, we firstly evaluate all controllers with different verhicle setups and the resultes are shown in Evaluation. Then we test our SAC controller for the generalization ability with unseen vehicle mass, tire friction and vehicle types. These results are shown in Generalization. We also provide the reference performances from the human driver and the test result with a rough and easy-to-access reference trajectory, which are listed in AppTest.
Benchmark Table
