Stanford Engineering and Toyota Research Institute Achieve World’s First Autonomous Tandem Drift
Teams Collaborate on AI-Powered System to Make Driving Safer
By Trey Weber, Daiki Mori, Nicholas Broadbent, Michael Thompson, John Subosits, Takao Kobayashi, Mason Llewellyn, Avinash Balachandran, J. Christian Gerdes
Leveraging the latest AI technology, Stanford Engineering and Toyota Research Institute are working to make driving safer for all. By automating a driving style used in motorsports called ‘drifting’ — where a driver deliberately spins the rear wheels to break traction — the teams have unlocked new possibilities for future safety systems.
Introduction
To understand how automated vehicles (AVs) can navigate safety-critical scenarios, we explore extreme vehicle dynamics and control. Researchers at both institutions have demonstrated that an AV can be controlled precisely even when operating with the rear tires completely sliding at the friction limits [1–4]. This capability can, in the future, help automobiles recover from a spin-out or maintain control on a slippery winter road.
On October 21, 2015, the Dynamic Design Lab at Stanford Engineering unveiled MARTY, an electric autonomous drifting DeLorean that could perform steady circular drifts. In 2022, TRI researchers demonstrated single-car drifting on a race track. Recently, we challenged ourselves to explore the limits of these ideas by seeking out the most challenging motion planning and vehicle control problem we could imagine: tandem drifting.
Tandem drifting, popularized in the US by Formula Drift, is a competition where two skilled drivers drift simultaneously. While the lead vehicle focuses on tracking an ideal path, or line, at a high angle, the chase vehicle has a much more complicated task. It must chase the lead car, maintaining close proximity as they drift through the course together, all while avoiding any collisions. This is a valuable problem to study for autonomy systems because it requires them to balance multiple conflicting objectives, such as avoiding collisions, staying on the road, and reacting to other vehicles in real time.
Hardware
To demonstrate autonomous tandem drifting, we tested two autonomous vehicles that can execute these maneuvers. We introduce “Keisuke,” the TRI vehicle and “Takumi,” the Stanford vehicle. Each automated and modified GR Supra has increased horsepower (700 for Keisuke and 520 for Takumi) and meets the safety standards of Formula Drift.
Keisuke and Takumi are equipped with computers and sensors that allow them to control their steering, throttle, and brakes while also sensing their motion (e.g., position, velocity, and rotation rate). However, crucial for the precise coordination of the two vehicles, Keisuke and Takumi can communicate by sharing a dedicated WiFi network, which allows them to exchange information, such as where they are relative to each other and their current planned trajectories.
Motion Planning and Control
To drift, Takumi and Keisuke must continually plan their steering, throttle, and brake commands and the trajectory they intend to follow. They do this using a technique called Nonlinear Model Predictive Control (NMPC). Each vehicle starts with a set of instructions or objectives, represented mathematically as a cost function and a set of rules or constraints they must obey. Then, each car solves an optimization problem to decide what the steering, throttle, and brake commands should be over the next 2–3 seconds to best meet their objectives. They process this problem up to 50 times each second to respond to the rapidly changing conditions.
Keisuke’s NMPC objective as the lead vehicle is to sustain a drift while following a desired path, subject to constraints like the laws of physics and the vehicle’s hardware capabilities — like its maximum steering angle. While that might seem straightforward, Keisuke must do this reliably every single time because if it enters a spin, Takumi will have nowhere to go to avoid a crash.
As the chase vehicle, Takumi must drift alongside Keisuke while proactively avoiding a collision. The most challenging point comes when Keisuke transitions or changes direction. Takumi must drop back to give Keisuke room and then accelerate to catch up, all while maintaining safety. To assist Takumi, Keisuke communicates its current position and planned maneuvers over the WiFi connection.
Leveraging AI
At the heart of the NMPC algorithm is a model describing how steering, throttle, and brake commands generate the tire forces that ultimately determine the vehicle motion. The accuracy with which this underlying model predicts real-world behavior determines the performance that can be achieved with this approach. However, the real world is a challenge to model.
How do you model a tire as it reaches over 212 degrees F (or 100 degrees C) and literally turns into white smoke as it drives? Takumi solves this problem with a physics-informed neural network that leverages the benefits of data-driven methods while blending in physics knowledge. Using data from previous tests, we train a neural network model of how the front tires react to combined cornering and brake forces. Because friction between the tire and the road is temperature-dependent, the rear tire forces depend heavily on tire temperature. The rear tire model captures this dependence with a physical model that uses tire temperature measurements to model friction changes. The physics-informed neural network demonstrates greater accuracy and robustness to ambient temperature changes than our traditional physics-based models. By leveraging AI, Takumi can learn and improve from every trip to the track.
A model that is constantly changing, however, poses challenges for a safety-critical application like tandem drifting. To capture the performance benefits of AI, we first needed to build confidence. We did this by collecting a baseline training data set and validating that our training process generated models successfully using this data. We then used experiments over several months and a small set of simulation data to expand this data and produce more reliable models. Finally, since the track temperatures at night when we performed the test were significantly lower than during the day when we collected the bulk of the data, we formulated a strategy to safely augment the training data set with data taken at night.
The result was exactly what we had hoped for — a spectacular yet safe, autonomous tandem drift. While your future car may not drift its way to the supermarket, it may very well incorporate some of the motion planning, control, and AI techniques developed in this project to keep you safe when road conditions or circumstances turn difficult or dangerous.
- T. P. Weber and J. C. Gerdes, “Modeling and Control for Dynamic Drifting Trajectories,” in IEEE Transactions on Intelligent Vehicles, vol. 9, no. 2, pp. 3731–3741, Feb. 2024, doi: 10.1109/TIV.2023.3340918.
- Goh JY, Thompson M, Dallas J, Balachandran A. Beyond the stable handling limits: nonlinear model predictive control for highly transient autonomous drifting. Vehicle System Dynamics. 2024 Feb 29:1–24.
- T. Kobayashi, T. P. Weber, D. Mori and J. C. Gerdes, “Trajectory Planning Using Tire Thermodynamics for Automated Drifting,” 35th IEEE Intelligent Vehicles Symposium (IV2024), June 2–5, 2024, Jeju Island, Korea. https://ieeexplore.ieee.org/abstract/document/10588753
- N. D. Broadbent, T. P. Weber, D. Mori and J. C. Gerdes, “Neural Network Tire Force Modeling for Automated Drifting,” 16th International Symposium on Advanced Vehicle Control (AVEC24), September 2–6, 2024, Milan, Italy. arXiv:2407.13760 [eess.SY]
