Learning to Drift in Extreme Turning with Active Exploration and Gaussian Process Based MPC

Guoqiang Wu, Cheng Hu, Wangjia Weng, Zhouheng Li, Yang Fu, Lei Xie, Hongye Su

TL;DR

This paper tackles the challenge of precise vehicle control during extreme cornering (drifting) by proposing a novel Model Predictive Control (MPC) framework enhanced with Gaussian Process Regression (GPR). The key innovation is using GPR not only to correct vehicle model inaccuracies but also to actively explore different drifting speeds, significantly improving trajectory tracking accuracy in challenging maneuvers.

What Problem Does This Paper Address?

Autonomous driving systems often struggle with extreme maneuvers like high-speed drifting. Conventional controllers are designed for linear tire-slip regimes and fail when tires experience significant slip angles, leading to poor control precision and potential instability. While drifting can actually improve cornering performance, it introduces significant vehicle model mismatches due to these large slip angles. Existing Model Predictive Control (MPC) approaches for drifting are sensitive to these model inaccuracies, impacting both the calculation of the drift equilibrium point and the optimization process itself. This paper aims to bridge this gap by developing a robust drifting controller that can accurately handle model uncertainties and actively improve its understanding of the vehicle's behavior in drift conditions.

Methods

The core of the proposed solution is the AEDGPR-MPC (Active Exploration with Double-layer GPR and MPC) system. It's a multi-stage approach:

  1. Minimum Time Planner: This module first generates an optimal, minimum-time trajectory for the vehicle, identifying segments where drifting is necessary.
  2. Active Explorer: This is a novel component. It takes the planned trajectory and uses the uncertainty estimates from GPR to generate a set of candidate cornering speeds. It then intelligently selects a speed (Vep) to explore, balancing the need to cover new ground in the GPR's knowledge space with the desire to stay close to the reference trajectory.
  3. Learning-Based State Planner: This module uses a GPR-corrected vehicle model to determine the equilibrium drift states (like sideslip angle and yaw rate) corresponding to the explored speed (Vep).
  4. Learning-Based MPC Controller: This is the final control layer. It uses the GPR-corrected model to track the reference states planned by the state planner, executing the drift maneuver.

The "double-layer GPR" refers to its application in both the equilibrium point calculation and the MPC optimization, ensuring consistent model correction. The "active exploration" is the key differentiator, using GPR's variance to guide the search for optimal drifting speeds.

Datasets

The paper does not explicitly mention or introduce new datasets for training. Instead, it relies on data collected during the experiments:

  • Simulation: Data was collected using the Simulink-Carsim platform.
  • Real-world Experiments: Data was collected from a 1:10 scale RC car.

The GPR model is trained and updated online based on the data gathered during these experimental runs.

Key Insights and Ideas

  • GPR for Model Correction is Crucial: The paper strongly demonstrates that GPR significantly improves control accuracy by compensating for model mismatches inherent in drifting.
  • Active Exploration is Key to Performance Gains: Simply using GPR isn't enough. Actively exploring different speeds using GPR's uncertainty estimates leads to substantial further improvements, suggesting that GPR's effectiveness is highly dependent on the richness and diversity of its training data.
  • Balancing Exploration and Exploitation: The active exploration algorithm cleverly balances exploring new, uncertain regions of the vehicle's state space with maintaining stable control by staying close to the reference trajectory. This is achieved by trading off distance to the reference and GPR variance.
  • Drifting as a Control Strategy: The work reinforces the idea that drifting, when properly controlled, is not just a stylistic element but a viable and effective strategy for navigating extreme cornering scenarios.

Key Results

The paper reports significant improvements in both simulation and real-world experiments:

Simulation (Simulink-Carsim): * Average lateral error reduced by 52.8% with GPR compared to non-GPR MPC. * Further reduction of 27.1% in average lateral error after active exploration. * Velocity tracking RMSE decreased by 10.6% with exploration.

Real-world RC Car Experiment: * Average lateral error reduced by 36.7% with GPR compared to non-GPR MPC. * Further reduction of 29.0% in average lateral error after active exploration. * Velocity tracking RMSE improved by 7.2% with exploration.

Limitations and Future Work

Limitations: * The paper focuses on a specific type of extreme cornering (drifting). Its applicability to other extreme maneuvers might require further investigation. * The computational cost of GPR and MPC, especially in real-time applications, could be a factor for more complex scenarios or hardware.

Future Work: * The authors plan to scale this algorithm to full-sized vehicles. * They aim to evaluate its performance across a wider range of extreme conditions.

Summary updated

Abstract

Extreme cornering in racing often leads to large sideslip angles, presenting a significant challenge for vehicle control. Conventional vehicle controllers struggle to manage this scenario, necessitating the use of a drifting controller. However, the large sideslip angle in drift conditions introduces model mismatch, which in turn affects control precision. To address this issue, we propose a model correction drift controller that integrates Model Predictive Control (MPC) with Gaussian Process Regression (GPR). GPR is employed to correct vehicle model mismatches during both drift equilibrium solving and the MPC optimization process. Additionally, the variance from GPR is utilized to actively explore different cornering drifting velocities, aiming to minimize trajectory tracking errors. The proposed algorithm is validated through simulations on the Simulink-Carsim platform and experiments with a 1:10 scale RC vehicle. In the simulation, the average lateral error with GPR is reduced by 52.8% compared to the non-GPR case. Incorporating exploration further decreases this error by 27.1%. The velocity tracking Root Mean Square Error (RMSE) also decreases by 10.6% with exploration. In the RC car experiment, the average lateral error with GPR is 36.7% lower, and exploration further leads to a 29.0% reduction. Moreover, the velocity tracking RMSE decreases by 7.2% with the inclusion of exploration.
BibTeX
@misc{wu2024learning,
  author = {Guoqiang Wu and Cheng Hu and Wangjia Weng and Zhouheng Li and Yang Fu and Lei Xie and Hongye Su},
  title = {{Learning to Drift in Extreme Turning with Active Exploration and Gaussian Process Based MPC}},
  year = {2024},
  eprint = {2410.05740},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2410.05740},
}
Figure