Driving Innovation: The Science of Torque Vectoring Systems

In the era of widespread adoption of AI models, the data science community exhibits a rich diversity of backgrounds, fostering an environment conducive to innovation. Collaborative efforts stemming from this diversity continue to propel us toward the envisioned future of a decade or two ago. However, innovation often follows incremental patterns and may require decades before yielding tangible commercial benefits.

Among industries adept at translating ideas into tangible products, automotive manufacturers stand out, boasting a deep familiarity with research and development processes and advanced algorithms. This proficiency is evident in the innovation pipeline for super and hypercars, drawing inspiration from advancements in rocket science. Leveraging extensive simulations and field tests, these manufacturers consistently push the boundaries of performance and safety in vehicle design.

Major manufacturers have achieved breakthroughs with torque vectoring systems by capitalizing on established development pipelines and harnessing the capabilities of reinforcement learning. The late 2010s witnessed a surge in commercial vehicles featuring these systems, attributed partly to their association with electric and hybrid vehicles and partly to the appeal of enhanced safety.

While the full extent of their influence on driver experience remains to be realized, these systems have already demonstrated remarkable efficacy, particularly in high-performance vehicles. So, what exactly is a torque vectoring system, and how is this innovation shaping the landscape of automotive manufacturing?

What is Torque Vectoring

In Torque Vectoring Systems (TVS), torque denotes a rotational force applied to a vehicle’s differentials or generated by independent electric motors on its wheels. These systems aim to regulate this rotational force to enhance speed, stability, handling, balance, and other performance aspects.

A compelling illustration of the protracted nature of innovation is evident in the evolution of torque vectoring systems, with the first cars equipped with Torque Vectoring Systems (TVS) debuting in the 1990s, pioneered by Mitsubishi. Initially known as Active Yaw Control, Mitsubishi integrated this technology into vehicles like the 1996 Mitsubishi Lancer Evolution IV GSR models.

The initial implementation relied on steering, throttle inputs, and g-force sensors to mitigate torque steer, a phenomenon where engine torque influences steering behavior. Initially geared towards racing applications, these systems have since evolved to cater to broader audiences, with companies like Rimac continuing to employ them for racing. In contrast, others emphasize safety and enhanced handling features for everyday drivers. Furthermore, advancements driven by the pursuit of eco-friendly transportation have led to the utilization of independent electric motors in torque vectoring systems, further expanding their applicability.

Commercial Applications

Ricardo (Honorable Mention)

In 2006, Ricardo, a renowned engineering consultancy with over a century of experience, introduced the concept of torque vectoring. Ricardo rolled out its initial torque vectoring project centered around the Audi A6 4.2L, intending to license its technology to interested automakers. Unlike Mitsubishi’s Active Yaw Control system and similar counterparts, Ricardo’s torque vectoring approach prioritized safety and stability, showcasing a different aspect of the technology.

Limited information is currently available on Ricardo’s involvement in assisting manufacturers in developing these systems for commercially available vehicles. However, they have publicly provided engineering support to Chinese students building cars for the global multi-event Formula SAE race in 2019.

Rimac

When discussing electric vehicles, Tesla typically takes the spotlight regarding market dominance. However, Rimac stands out as a company dedicated to crafting high-performance electric hypercars, where they excel. Just ask Elon… (X post)

Rimac burst onto the scene in January 2016 with their inaugural concept car, the Concept_One prototype. This marked the debut of their torque vectoring system (TVS) in an all-wheel-drive fully electric vehicle. Spearheaded by Tomislav Šimunić, Rimac’s TVS development pushed the boundaries by employing individual motors on each wheel. Moreover, they implemented reinforcement algorithms utilizing data from various sensors, including accelerometers, gyroscopes, steering angle, and wheel speed sensors. This data is then used to predict torque distribution to optimize vehicle stability.

Currently, Rimac boasts one of the fastest and safest hypercars on the market. Their innovation has led to unprecedented control, resulting in their car achieving the title of the fastest production car ever in 21 acceleration and deceleration categories. Furthermore, their TVS is so reliable that their drivers can cruise hands-free at speeds exceeding 230 mph while enjoying gelato!

Toyota

In 2018, Toyota introduced their Dynamic Torque Vectoring AWD and Electric 4WD systems, expanding the scope of this technology beyond on-road safety and control to encompass off-road safety features and enhanced fuel efficiency.

The video above showcases Toyota’s emphasis on “proactive torque transfer,” allowing for seamless adjustment of torque distribution when one or more wheels lose traction, enabling drivers to navigate various terrains more effectively.

Furthermore, Toyota optimizes torque distribution to enhance fuel-efficiency, shifting from AWD to 2WD when conditions permit, directing torque predominantly to the front wheels and thus cutting fuel consumption. Their limited content suggests the employment of a Dynamic Control System designed for stability, taking into account challenging driving scenarios. Moreover, the system can evaluate tire traction levels. These functionalities resonate with the research findings discussed below, showcasing the seamless transition from research concepts to practical implementation in the automotive industry.

Research Review

Conceptual Models — Bicycle, Vehicle, Tire Model

Following Occam’s razor principle, simplicity reigns supreme when initiating modeling endeavors, with complexity introduced only when its benefits are worthwhile. This philosophy holds for TVS as well. Early TVS adopted a bicycle model, allowing researchers to streamline vehicle dynamics assumptions by confining movements to two wheels. This simplified framework facilitated the study of lateral vehicle dynamics during various events like braking, acceleration, and cornering under typical conditions.

Key inputs into the bicycle model include yaw rate, lateral acceleration, steering angle, and velocity. Yaw rate denotes the vehicle’s rotation speed around its vertical axis, aiding in determining its orientation during maneuvers like cornering. Lateral acceleration captures sideways movements perpendicular to the vehicle’s direction of travel. Steering angle refers to the front wheel angle, typically in relation to the rear wheel positioning, while velocity represents the vehicle’s speed.

For more details, check out this article explaining a simple bicycle model.

Simplicity in modeling often comes with assumptions that may not hold, especially in challenging driving scenarios. Extreme conditions like rain, snow, or rough off-road terrain expose the limitations of the traditional bicycle model. The assumption of symmetry between the left and right wheels hinders the accurate analysis of lateral dynamics across all four wheels. To address this, researchers upgraded to a more comprehensive 4-wheel vehicle model.

One such model, proposed in the 2023 journal article Deep Reinforcement Learning-Based Torque Vectoring Control Considering Economy & Safety² by Hui Deng, Youqun Zhao, Fen Lin, and Qiuwei Wang, boasts seven degrees of freedom. This expanded model captures all four wheels’ lateral, longitudinal, rotational, tilting, leaning, and spinning movements. However, it still falls short in capturing driving dynamics such as suspension behavior and tire-road interactions. Although suspension dynamics may be indirectly represented within the model, a separate tire model is necessary to capture tire-road interactions accurately.

Tire models typically include parameters like tire force and slip angles. In the journal article A Torque Vectoring Control for Enhancing Vehicle Performance in Drifting³ by Michele Vignati, Edoardo Sabbioni, and Federico Cheli, researchers describe a simple tire model that incorporates lateral and longitudinal dynamics, slip angles, and slippage. Like the seven degrees of freedom model indirectly addresses suspension dynamics, the tire model considers slippage influenced by latent variables such as surface texture, conditions, and angles. This comprehensive approach allows researchers to analyze vehicle dynamics across various conditions and terrains, enhancing the torque vectoring system’s adaptability and performance.

Control Schemes/Implementations

In Torque Vectoring Systems (TVS) development, a significant portion of research revolves around employing Reinforcement Learning techniques, leveraging real-time data collected from various measurement devices. These solutions are predominantly implemented in all-wheel drive, four-wheel drive hybrid, and electric vehicles. Typically, the training of these systems takes place within simulated environments. However, in certain instances, researchers conduct road tests and utilize live test outcomes to validate simulation results.

Torque Vectoring System Testing

While specific details about testing TVS are often limited, researchers generally have three options: entirely computer-based simulations, hardware-in-the-loop testing combining simulation with real parts, and full-scale vehicle road tests. Computer-based simulations offer cost-effective evaluation of TVS with precise control over various driving conditions. Hardware-in-the-loop testing balances simulation and road tests, providing insights aligned with real-world performance. Although the most expensive, road tests offer the most reliable results by implementing TVS fully, enabling the most accurate evaluation of model performance.

In the study Self-adaptive Torque Vectoring Controller Using Reinforcement Learning⁴ by Shayan Taherian, Sampo Kuutti, Marco Visca, and Saber Fallah, computer-based simulations were utilized to test their RL algorithm across diverse environments. The researchers employed an adaptive algorithm based on Deep Deterministic Policy Gradient to update TVS control parameters as it learns from the environment. Testing across various velocity and friction environments allowed validation of their adaptive algorithm, particularly in high-speed/low-velocity scenarios.

Similarly, in the research Torque Optimization Control for Electric Vehicles with Four In-Wheel Motors Equipped with Regenerative Braking System⁵ by Wei Xu, Hong Chen, Haiyan Zhao, and Bingtao Ren, real-time Simulink simulations were leveraged to validate their model predictive controller. This controller predicts future system states and optimizes for multiple energy efficiency and safety objectives. Tests were conducted using an electric vehicle model with a 120 km/h velocity and an “adhesion coefficient” of .85, measuring tire friction between tire and ground.

Powertrain & Fuel Types

The specific problem addressed by torque vectoring systems heavily influences the powertrain choice and fuel types utilized in their implementations. Traditional gasoline and diesel vehicles necessitate mechanical implementations due to a single torque source, leading to complex mechanical differentials, challenges in power distribution, and overall less efficiency compared to hybrid and electric vehicle setups, which enable independent motor implementations.

Similarly, All-Wheel Drive (AWD) and Four-Wheel Drive (4WD) powertrains offer superior advantages over other options. These configurations enhance stability, versatility, and energy efficiency. The work of Jeongmin Cho and Kunsoo Huh, in their study Torque Vectoring System Design for Hybrid Electric–All-Wheel Drive Vehicle⁶, underscores the stability and versatility benefits of AWD vehicles. By distributing power to all four wheels, AWD enhances the effectiveness of torque vectoring systems, granting increased control over each wheel. This translates to improved vehicle dynamics during cornering, acceleration, and deceleration.

Meta Control Algorithms

In Reinforcement Learning (RL), an agent decides actions based on the current situation to maximize a reward function. It learns which actions yield the highest rewards by assessing the outcomes of its actions. In torque vectoring algorithms, auto manufacturers define different reward functions based on the specific goals of their TVS, which vary depending on the type of vehicles used, such as sports cars versus mini-vans. These goals include increased stability, controllability, fuel economy, cornering performance, and optimizing various vehicle dynamics behaviors.

One example of a multi-objective implementation is outlined in the paper Development of Torque Vectoring Control Algorithm for Front Wheel Driven Dual Motor System and Evaluation of Vehicle Dynamics Performance⁷ by Jae-Young Park, Seung-Jin Heo, and Daeoh Kang. Their algorithm features a supervisory controller that selects between an “agile” mode and a “safe” mode based on vehicle and driving conditions. Agile mode prioritizes driver controllability, while safe mode focuses on stability. Controllability refers to the synchronization between driver steering input and vehicle direction, while stability relates to the vehicle’s resistance to deviations or disturbances. Inputs such as vehicle speed, steering angle, yaw rate, lateral acceleration, and wheel slip guide the supervisory controller’s decision-making process.

Another example is presented in the paper Deep Reinforcement Learning-Based Torque Vectoring Control Considering Economy and Safety⁸ by Hui Deng, Youqun Zhao, Fen Lin, and Qiuwei Wang. Here, the researchers optimize for safety, economy, and driver load. Economy involves assessing motor efficiency and energy savings compared to vehicles without torque vectoring systems. Safety focuses on vehicle stability, while driver load considers the physical and cognitive load based on the correlation between the driver’s steering wheel angle and the vehicle’s longitudinal acceleration.

The algorithm includes an active safety controller and a torque allocation layer. The active safety controller calculates total longitudinal acceleration and additional “yaw moment” values based on driving conditions, applying safety and driver load penalties. These values are then fed to the torque allocation layer, which distributes torque to all four wheels. Additionally, the researchers implement a tire model using a Fibonacci tree optimization algorithm and a vehicle model with 7 degrees of freedom to enhance system performance.

The Torque Vectoring Algorithm of Electronic-Four-Wheel Drive Vehicles for Enhancement of Cornering Performance⁹, research done by Park et al., aims to enhance the driver’s experience through improved smoothness and convergence. Their TVS, designed for commercial application, underwent rigorous validation through real vehicle experiments conducted under various conditions.

The researchers implemented a smooth sliding controller, GPS, and sensor inputs to generate desired yaw moment values. These values are then utilized by the torque allocation algorithm, which employs a daisy-chaining allocation method to distribute torque efficiently. The goal is to optimize the convergence from current yaw moment values to target values, thereby reducing the instability during this transition.

Conclusion

The evolution of torque vectoring systems underscores the transformative power of innovation within the automotive industry. From pioneering beginnings to cutting-edge applications, these systems have revolutionized vehicle dynamics, speed, stability, and handling. As automotive manufacturers continue to push the boundaries of technology and leverage advancements in advanced algorithms, the future holds promise for further improvements in safety, performance, and driver experience. However, hurdles remain in integrating this technology into internal combustion engine vehicles, and ongoing research and development efforts are required to overcome them. With continued dedication to innovation, torque vectoring systems are poised to shape the future landscape of automotive manufacturing, driving us toward a new era of mobility.

Research Reference

1. Martin Dendaluce Jahnke et al., “Efficient Neural Network Implementations on Parallel Embedded Platforms Applied to Real-Time Torque-Vectoring Optimization Using Predictions for Multi-Motor Electric Vehicles,” Electronics 8, no. 2 (February 22, 2019): 250

   
   

2. Huifan Deng et al., “Deep Reinforcement Learning-Based Torque Vectoring Control Considering Economy and Safety,” Machines 11, no. 4 (April 6, 2023): 459

   
   

3. Michele Vignati, Edoardo Sabbioni, and Federico Cheli, “A Torque Vectoring Control for Enhancing Vehicle Performance in Drifting,” Electronics 7, no. 12 (December 5, 2018): 394

   
   

4. Shayan Taherian et al., “Self-Adaptive Torque Vectoring Controller Using Reinforcement Learning,” 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), September 19, 2021

   
   

5. Wei Xu et al., “Torque Optimization Control for Electric Vehicles with Four In-Wheel Motors Equipped with Regenerative Braking System,” Mechatronics 57 (February 2019): 95–108

   
   

6. Jeongmin Cho and Kunsoo Huh, “Torque Vectoring System Design for Hybrid Electric–All Wheel Drive Vehicle,” Proceedings of the Institution of Mechanical Engineers, Part D: Journal of Automobile Engineering 234, no. 10–11 (April 7, 2020): 2680–92

   
   

7. Jae-Young Park, Seung-Jin Heo, and Dae-Oh Kang, “Development of Torque Vectoring Control Algorithm for Front Wheel Driven Dual Motor System and Evaluation of Vehicle Dynamics Performance,” International Journal of Automotive Technology 21, no. 5 (October 2020): 1283–91

   
   

8. Huifan Deng et al., “Deep Reinforcement Learning-Based Torque Vectoring Control Considering Economy and Safety,” Machines 11, no. 4 (April 6, 2023): 459

   
   

9. Giseo Park et al., “Torque Vectoring Algorithm of Electronic-Four-Wheel Drive Vehicles for Enhancement of Cornering Performance,” IEEE Transactions on Vehicular Technology 69, no. 4 (April 2020): 3668–79