Osegi
Science·1 min read

AI teaches a 5-tonne John Deere 332G to balance on two wheels

Researchers got a 5-tonne John Deere 332G to lift its rear wheels and drive on two, steered by an AI trained only in simulation on flat ground.

A heavy John Deere 332G compact track-style loader balancing on its two front wheels with its rear wheels lifted off the ground.
Image: 미래를 보는 창 - 전자신문

A 5-tonne construction machine standing on two wheels sounds like a bad day on a building site. Here it is the point. Researchers, including a team from Switzerland's ETH Zurich, have used AI to make a John Deere 332G balance and drive on its front wheels alone.

The machine starts by rocking back and forth to build up momentum. That swing lifts the rear wheels off the ground. Once it is up, the AI keeps it balanced on the two front wheels while it drives forward and backward, spins on the spot and moves its bucket. Then it lowers itself back onto all four wheels, gently.

How was the AI trained?

The team used reinforcement learning, a method where software learns by trial and error and gets rewarded for good results. All of that practice happened in a virtual environment, and the training terrain was flat. The trained controller was then put on the real machine.

That is where it got interesting. In real tests, the machine crossed uneven ground with bumps 75 cm high while staying on two wheels, even though it had never seen such terrain in training.

What is it for?

The stunt is not the goal in itself. According to the researchers, the work points to practical uses. One is a heavy machine that climbs onto a truck by itself, with no ramp. Another is turning with less damage to the ground underneath.

For now this is a research demonstration, and the source gives no timeline for commercial use. It does show that learning-based control, trained in a simulator, can handle a machine this heavy and awkward in the real world.

Apps worth a look

SUMO: the number puzzle

Tap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.

Tags:john deereeth zurichreinforcement learningroboticsexcavator332g

More from this day

See all →

More in: Science

See all →