Tuning the feedback controller gains is a simple way to improve autonomous driving performance

TL;DR

Autonomous driving systems combine NN-based planning with classical feedback controllers, but controller gains are often not re-tuned when NN components change, potentially degrading closed-loop performance. The authors show that a simple, manual retuning of the longitudinal PID gains in the TCP algorithm can meaningfully improve driving performance in CARLA, boosting the Driving Score from $73.21$ to $77.38$ across 16 scenarios, especially in rain and night conditions. The approach uses a two-branch TCP architecture with situation-based fusion and demonstrates that a more aggressive PD response (higher $k_p$, lower $k_i$) and modest brake adjustments can reduce crashes and improve robustness. The work advocates closer collaboration between control theory and machine learning in ADS design and suggests avenues for further improvement via automatic tuning or model-predictive control (MPC) to exploit NN-generated waypoints. While preliminary and based on manual tuning, the results indicate a practical, scalable path to enhanced performance without retraining neural components.

Abstract

Typical autonomous driving systems are a combination of machine learning algorithms (often involving neural networks) and classical feedback controllers. Whilst significant progress has been made in recent years on the neural network side of these systems, only limited progress has been made on the feedback controller side. Often, the feedback control gains are simply passed from paper to paper with little re-tuning taking place, even though the changes to the neural networks can alter the vehicle's closed loop dynamics. The aim of this paper is to highlight the limitations of this approach; it is shown that re-tuning the feedback controller can be a simple way to improve autonomous driving performance. To demonstrate this, the PID gains of the longitudinal controller in the TCP autonomous vehicle algorithm are tuned. This causes the driving score in CARLA to increase from 73.21 to 77.38, with the results averaged over 16 driving scenarios. Moreover, it was observed that the performance benefits were most apparent during challenging driving scenarios, such as during rain or night time, as the tuned controller led to a more assertive driving style. These results demonstrate the value of developing both the neural network and feedback control policies of autonomous driving systems simultaneously, as this can be a simple and methodical way to improve autonomous driving system performance and robustness.

Figures (5)

• Figure 1: Schematic of the TCP algorithm from tcp.
• Figure 2: Images from the drivers perspective of the vehicle in four different weather scenarios.
• Figure 3: Comparison between the original driving scores for the original TCP algorithm tcp and that with the tuned PID gains for the longitudinal controller, referred to as TCP-tuned.
• Figure 4: Four examples of crashes recorded with the TCP tcp algorithm in CARLA (ego vehicle highlighted by red box).
• Figure 5: Instances of the TCP algorithm with tuned PID gains avoiding crash events. In both cases, the ego vehicle is highlighted by a red bounding box. The bicycle in b) is highlighted by a blue bounding box.