Robotics Weekly

A 5-tonne John Deere loader balances on two wheels by itself

A learned policy from Marco Hutter's lab pulls off a stunt only a handful of elite operators can do, on an operational construction machine.

A team from ETH Zurich's legged robotics lab taught a policy to tip a John Deere 332G skid-steer loader onto two wheels and keep it there, no human at the controls. The machine is an operational 5-tonne skid-steer. The policy swings it up, balances while driving over rough ground, traces set trajectories, spins, swings the bucket back and forth, then sets it back down on four wheels.

As @lukas_m_ziegler puts it, tipping a skid-steer up and driving it like that is a stunt only a handful of elite operators can do, and they do it for world records.

Lukas Ziegler
@lukas_m_ziegler
X
Only a handful of elite operators can do it, and they do it for world records.
Nov 18, 2026 · View on X
Lukas Ziegler
@lukas_m_ziegler
X
the real achievement is modeling the hydraulics faithfully enough that the learned behavior transfers to the physical loader
Nov 18, 2026 · View on X

Why hydraulics make this hard

Two things stand in the way. The first is the hydraulics. These machines move with fluid power, which brings flexible elements and nonlinear delays that are nasty to model and control. Most of the sim-to-real (trained in simulation, run on the real machine) work people have seen lately runs on electric actuators with clean, fast responses. Hydraulics do not behave like that.

The second is safety. A 5-tonne machine at the edge of its balance limits is far too dangerous to learn on by trial and error. You cannot let it fall over a thousand times on a real site.

Reinforcement learning in simulation deals with both at once. The policy trains across countless simulated runs where crashing costs nothing. The real achievement, as Ziegler describes it, is modeling the hydraulics faithfully enough that the learned behavior transfers to the physical loader.

Same recipe, bigger machine

This is the sim-to-real approach that has been driving a lot of legged robotics, pointed at construction equipment instead. It works on a machine that weighs as much as a small truck, which is the part worth paying attention to. The gap between a quadruped learning to walk in sim and a hydraulic loader balancing on two wheels in the real world is mostly a modeling gap, and this closes it on a commercial machine.

The work comes out of Marco Hutter's lab at ETH Zurich, with industry partners including John Deere, and is published in Science Robotics.

One thing the source does not say is how reliably the loader does this, how many runs were attempted, or what the recovery behavior looks like if the balance goes wrong. The video shows the full sequence including the controlled return to four wheels. Everything beyond that is not in the public post yet.

Sound like you were at the demo

One email a day. Under three minutes. The demos, the deployments and the part the press release left out.

Get the next one by email

Robotics every day from the people building it. The demos, the deployments and the arguments worth your time. Every claim links back to the engineer.