Logo of Acerta
Contact us

Case Study

Estimating road friction in milliseconds

Objectives

Enable real-time estimation of road surface friction to improve vehicle safety and performance, using efficient machine learning models deployable on embedded systems.

Challenge

Normal driving data lacks clear road grip signals, and the model had to be under 5 MB to fit the vehicle’s hardware.

Key Results

  • Detected road surface changes in less than 1 second
  • Shrunk model size by 97% for real-time use on the vehicle
  • Exceeded OEM’s accuracy target by 32%

>1 sec

Detection time

97%

Reduction in model size

32%

Higher accuracy than target

>1 sec

Detection time

97%

Reduction in model size

32%

Higher accuracy than target

Background

A North American OEM sought to improve vehicle dynamics by accurately estimating the maximum available friction (surface μ) between tires and road surfaces in real time. Identifying optimal wheel slip ratios is essential not only for anti-lock braking systems (ABS), but also for torque vectoring, transmission shift timing, and other key vehicle functions. However, gathering sufficient friction data typically requires high-excitation driving events like sudden braking—impractical in many real-world conditions, especially for electric vehicles (EVs) utilizing regenerative braking. To address this, the OEM partnered with Acerta to use machine learning for estimating road friction based solely on low-excitation vehicle data, enabling real-time insights without requiring extreme driving maneuvers.

Problem

Low-Excitation Data Limitations

The client used an EV that only used regenerative braking to decelerate while driving on a test track with an onboard data logger receiving signals from the vehicle’s standard array of sensors. The track incorporated a variety of road surfaces, including wet and dry asphalt, ice, and gravel.

in the absence of high-excitation events—such as rapid acceleration or braking—the feedback between the road and the vehicle has traditionally been insufficient to determine the maximum available surface friction, which is necessary for identifying optimal slip ratio.

Working with low-excitation data such as this has traditionally been difficult, since the indicators of surface friction are not usually recognized until just before a vehicle begins to slip. In usually recognized until just before a vehicle begins to slip. In other words, the model needed to recognize whether the electric vehicle was driving on wet asphalt or ice without any sudden changes in its speed.

Model Size Constraints

Initial machine learning models were around 100MB, far exceeding the 12MB hardware constraints of the OEM’s onboard systems.

Solution

Custom Model Development

The OEM generated a training data set, in which a driver initiated rapid braking on each surface of the test track to determine the coefficient of friction between the tires and the road surface. Acerta then evaluated the resulting data and built a machine learning model to identify the transition points between different road surfaces in the data set.

Next, Acerta’s data scientists selected several types of machine learning models—one incorporating proprietary algorithms— and tuned their parameters with feature engineering, continuously evaluating their performance on the training data. These were then tested on novel data for their speed and accuracy in detecting a change in road surface.

Edge Model Compression

Once optimal models were selected, Acerta’s engineers focused on compression techniques, successfully reducing the models and associated libraries from ~100MB to under 3MB—small enough to be deployed on a 12MB MicroAutoBox embedded system.

Results

  • <1 Second Detection Speed: The model identified road surface transitions in less than one second on average, enabling real-time friction estimation.
  • >97% Model Size Reduction: Machine learning models and libraries were compressed from approximately 100MB to under 3MB, meeting strict embedded hardware constraints.
  • Edge Deployment on 12MB System: Successfully integrated the compressed models into a 12MB MicroAutoBox, allowing real-time, onboard deployment without compromising performance.
  • Exceeded accuracy target by 32%: While the OEM was seeking an accuracy rate of at least 66%, Acerta’s model achieved 87% accuracy proportional to time.

We exceeded the accuracy target by 32%

"We achieved sub-second detection and exceeded the accuracy target by 32%, and we did it without any cameras or visual sensor data. It’s a powerful example of what happens when engineering precision and applied machine learning come together."
Sergey Strelnikov
VP of Engineering

Case Studies

Proven Impact