Infographic cover

📅 November 11, 2026 • Science Robotics
In Science Robotics on November 11, 2026 (Vol 11, Issue 120, article eaef0894) a peer-reviewed paper appeared under the title Heavy metal acrobatics: Learning dynamic control for skid-steer loaders. Per the paper, a policy trained by reinforcement learning in simulation and then transferred to real hardware autonomously performs acrobatic maneuvers on a John Deere 332G, an operational 5-tonne skid-steer loader. The work comes from the Robotic Systems Lab of ETH Zurich and Intel researchers, senior author Marco Hutter.

Per the paper, the policies swing the machine up onto two wheels and, while balancing on two wheels, drive over rough terrain, trace specified trajectories, spin around in a pirouette, and swing the bucket back and forth before landing safely back onto four wheels. Per the authors, such maneuvers were previously accomplished only by a small number of elite human operators in official world-record performances. That framing is the research team's own claim, and it recasts a 5-tonne hydraulic workhorse as an acrobat.

Why it is hard: per the paper, hydraulic actuation involves fluids, flexible elements and nonlinear delays, and machine size and weight make physical experiments exceedingly dangerous. Per the arXiv preprint, the test machine was retrofitted for the study — researchers intercepted the joystick signals and supplied control commands from a computer, wheel encoders were ex-works via the CAN bus, and two IMUs were added, one on the machine base and one on the arm. This is a research demonstration, not a shipping product, and no commercial deployment was announced.

Per DongA Science, the team logged about 1.5 hours of hydraulic response data, trained a simulator on four maneuvers, and the real machine ran on onboard sensors and a regular laptop. Per DongA Science, the practical direction is to let machines drive onto transport trucks without ramps, cross obstacles, regain balance in rough terrain, then add cameras and lidar and unify control behaviors. As automation of heavy machinery targets productivity in construction, mining and material handling, dangerous trials move into simulation and dynamic control goes beyond slow scripted motions.

Sources:
science.org/doi/10.1126/scirobotics.aef0894
arxiv.org/abs/2611.18365
tomshardware.com/tech-industry/robotics/autonomous-5-tonne-john-deere-skid-steer-loader-trained-to-perform-acrobatic-maneuvers-using-ai-researchers-discuss-reinforcement-learning-tech-in-a-paper-titled-heavy-metal-acrobatics
dongascience.com/en/news/80834
pubmed.ncbi.nlm.nih.gov/43013690

Key figures

IMPACT checklist

Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky

AI Automation