
A 5-ton(t) heavy construction vehicle, a skid-steer loader, performs stable stunt driving on just its two front wheels. Courtesy of Pierre-Jean Lorentz
A 5-ton(t) piece of heavy construction equipment has demonstrated unmanned stunt driving after training with artificial intelligence (AI). It successfully performed a variety of maneuvers—such as driving over a mound or spinning in place—while keeping the vehicle body stable on just its two front wheels.
A research team at the Robotic Systems Lab of ETH Zurich in Switzerland, together with Intel in the United States and others, has succeeded in controlling construction equipment autonomously using reinforcement learning–based AI to perform stunt driving. The team published the results on the 11th (local time) in the international journal Science Robotics.
Large construction machines such as excavators and loaders typically weigh several tons and are driven by hydraulic systems, which makes them difficult to automate. In hydraulic systems, the motion of internal fluids, friction, elasticity of components, and response delays to control commands interact in complex ways, making it hard to accurately predict actual movements. For safety reasons, it is also difficult to run repeated tests in real-world situations.
The team used reinforcement learning, a technique in which an AI repeatedly learns target behaviors through trial and error in a virtual environment. The experiment used John Deere’s 332G model, produced by the U.S. construction machinery manufacturer. It is a skid-steer loader weighing about 5 tons with four wheels, front and rear, that turns by exploiting differences in rotational speed between the left and right wheels and uses a front bucket to carry soil or materials.

A 5-ton(t) heavy construction vehicle, a skid-steer loader, performs stable stunt driving on just its two front wheels. Courtesy of Pierre-Jean Lorentz
The researchers measured basic physical properties of the machine, such as length and center of mass, and then fed various control commands into the actual equipment, collecting response data from the hydraulic drivetrain over about 1 hour and 30 minutes. Based on this data, they developed a neural network model to predict the machine’s motion and ran reinforcement learning in computer simulations.
To narrow the gap between the real machine and the virtual environment, they also applied a method that varied parameters such as the center of mass, inertia, and sensor error during training.
The team designed the system so that, even without demonstration data of a human driver performing stunt maneuvers, the AI could learn to lift the machine onto its two front wheels, maintain balance while driving, and then smoothly land back on four wheels.
In this process, four motions—lifting the machine onto two wheels, maintaining balance, controlling speed, and landing—were trained separately and then combined and executed on the real machine.
The loader moved back and forth to build up momentum, successfully lifting its rear wheels and balancing on its two front wheels. It maintained balance even when the arm connected to the bucket was moved up and down, shifting the center of mass. It also managed to spin in place and drive on just two wheels.

A 5-ton(t) heavy construction vehicle, a skid-steer loader, performs stable stunt driving on just its two front wheels. Courtesy of Pierre-Jean Lorentz
In repeated experiments with the real machine, when the bucket arm angle was 28 degrees and 31 degrees, it succeeded in all 11 and 12 trials, respectively, of lifting onto two wheels. At 34 degrees—outside the training range—it succeeded in 11 out of 14 trials, and the success rate dropped to 78.6%.
Although the AI had been trained only on flat terrain, it also succeeded in maintaining balance on two wheels while moving forward and backward over uneven ground that included a 75 cm-high mound.
The AI handled terrain changes using only onboard sensor data such as tilt and wheel speed, without any obstacle information from cameras or lidar that perceive surrounding terrain. When executing the pre-trained maneuvers on the real machine, it ran on an ordinary laptop equipped with an Intel Core i7 CPU and 16 GB of memory, without the need for a high-performance graphics processing unit (GPU).
This technology demonstrates that AI can autonomously perform complex maneuvers with heavy construction equipment that previously only a small number of highly skilled operators could execute. It suggests the possibility that construction machines could, for example, drive themselves onto transport trucks without separate ramps, travel over obstacles, and regain balance in rough terrain.
The team plans to develop this work into autonomous construction equipment technology capable of much more complex tasks by integrating cameras, lidar, and other devices that perceive the surrounding environment and by unifying multiple control behaviors.
– doi.org/10.1126/scirobotics.aef0894
Copyright ⓒ DongA Science. All rights reserved.