John Deere skid-steer learns acrobatic AI manoeuvres
Researchers trained a five-tonne John Deere 332G in simulation with reinforcement learning, then transferred the behaviour to the real loader. Videos show it balancing on two wheels and spinning.

Key takeaways
- Researchers trained a five-tonne John Deere 332G skid-steer loader to perform acrobatic manoeuvres autonomously.
- The system learned in simulation using reinforcement learning before being transferred to the real machine.
- Videos show the loader balancing on two wheels, spinning and controlling its bucket.
- The team presents the work as groundwork for automating heavy machinery in construction, mining and material handling.
In this article
What the loader can do
The videos show the 332G lifting itself onto two wheels and shifting its weight from side to side. It also turns on the spot and keeps its balance while moving the lift arm. According to Tom's Hardware, the loader can handle rough ground, spin and control its bucket before returning to a stable position.
The researchers say only a very small number of highly skilled human operators could perform similar stunts. That comparison is the team's claim, not an independent measurement.
How the machine learned

The team used reinforcement learning, a method in which software learns by trial and error from feedback. Training a machine of this size directly was not considered practical, so the work began in a simulation built around the 332G's controls. The simulation also used physics modelling and neural networks.
The learned behaviour was then moved to the real loader, a process known as sim-to-real transfer. The researchers present the result as evidence that simulation training can carry over to a heavy machine in the real world. The accuracy of the simulation is central to that claim.
Why heavy machinery is the target
The paper is presented as groundwork for autonomous control of heavy machines in demanding conditions. Tom's Hardware reports that construction, mining and material handling are among the industries that could benefit. The researchers hope automated control will eventually be safer, faster and more efficient than human operation.
That is a goal rather than a demonstrated outcome. The report does not include independent safety trials or a timetable for commercial use.
What it means for the sector
The demonstration suggests that automation may reach tasks that need quick balance and fine control, not only repeated driving. It also shows how simulation can reduce the cost and risk of teaching large machines new behaviour. For now, the work is a research result, not a product announcement.
Tom's Hardware argues that the videos point to wider scope for AI-driven automation in heavy equipment. The report does not address commercial deployment, staff training or regulatory approval.
Sources
Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.



