Self-Driving Car
Self-Driving Car
Embedded Files
BLUF - I implemented and tuned four different steering controllers to compare their performances on a car robot navigating a track.
Self-driving cars follow rigorous rules to ensure safety while traveling. Reference tracking algorithms ranging from PID to LQR provide the basis for which one day human drivers, who are prone to error, can be assisted with these techniques. By implementing various driving algorithms in Python and testing them in a simulated environment powered by WeBots, the performance of these methods were characterized and benchmarked.
Methods
Methods
Throughout these simulations, a bicycle dynamics model represented a Tesla Model 3 car that followed a reference trajectory. Throttle, steering angle, and deviation from the track were plotted over time. Controllers that were implemented for setting the steering angle include:
PID
PID
The proportional-integral-derivative controller is the industry standard for implementing a control system, but lacks long-horizon reasoning and accounting for uncertainty.
Pole-placement
Pole-placement
Stemming from classical control, pole placing generates stable controllers that minimize error between the car’s position and the reference trajectory.
LQR
LQR
Compared to PID, linear-quadratic regulators are more intuitive to tune, balancing state feedback and control inputs with weights.
EKF-SLAM
EKF-SLAM
Extended Kalman Filtering with simultaneous localization and mapping (EKF-SLAM) is the gold standard for modern robotic applications due to the ability to perform long-horizon reasoning and account for uncertainty.
Results
Results
Each controller outperformed specified benchmarks specified by the instructor after manual tuning of Kp/Ki/Kd values, poles of a 4th order controller, and Q/R matrices. The observability and controllability matrices were calculated and showed full rank for various speeds, indicating that the throttle and steering angle outputs were achievable throughout the runs.
Conclusion
Conclusion
With each successive algorithm, performance steadily improved. Programming self-driving cars to follow a reference trajectory closely and quickly proves the capacity for these technologies to one day augment the driving experience and guarantee safety. Future work involves tuning these controllers to further minimize deviation from the reference and applying them to other cars.
Page updated
Google Sites
Report abuse