Forget those tiny self-balancing Arduino robots.
Researchers just taught a neural network to balance a FIVE-TON construction loader on two wheels.
And that's only the beginning.
Meet Heavy Metal Acrobatics, a robotics research project published in Science Robotics by researchers from ETH Zurich and their collaborators.
The team used reinforcement learning to train an AI controller for a real John Deere 332G skid-steer loader.
Instead of manually programming every maneuver, they developed a physics-based simulation, trained control policies, and transferred them to the actual machine.
WHAT CAN THIS THING DO?
• Swing its entire 5-ton body onto its two front wheels.
• Maintain balance while driving forward and backward.
• Follow different trajectories while balanced.
• Drive across uneven and bumpy terrain.
• Spin around on two wheels.
• Move its hydraulic arms and bucket without losing balance.
• Return safely to all four wheels using a controlled landing.
And yes, the maneuvers are performed autonomously.
HOW DID THEY TRAIN IT?
The researchers first collected real machine data to model the loader's complex hydraulic behavior.
Then they used reinforcement learning in simulation to teach the controller how to balance and maneuver.
Finally, they transferred the learned policies to the real machine.
This approach is known as sim-to-real reinforcement learning.
THE INTERESTING PART
Balancing a tiny robot with two wheels is one thing.
Controlling five tonnes of steel, hydraulic actuators, momentum, and delayed mechanical responses is something completely different.
Some of these maneuvers were previously performed only by exceptionally skilled human operators.
Now a learned controller can execute them autonomously.
Of course, this is a controlled robotics research demonstration, not a construction vehicle ready to perform stunts on a public road or job site.
But it demonstrates how reinforcement learning could improve the dynamic control of heavy machinery in construction, mining, and industrial applications.
And I have to admit, watching a neural network teach a five-ton loader to do wheelies is considerably more entertaining than another tiny robot balancing on an Arduino.
RESEARCH & DEMOS
Official project, videos and supplementary experiments:
https://sites.google.com/leggedrobotics.com/heavy-metal-acrobatics
Research paper:
https://arxiv.org/abs/2611.18365
Published in Science Robotics, November 11, 2026.
What's more impressive: humanoid robots doing backflips or a five-ton construction machine balancing on two wheels?
#Robotics #ReinforcementLearning #ArtificialIntelligence #HeavyMachinery #Shorts
YouTube / ABVX
