A reinforcement-learning system has taught a five-ton construction loader to balance and move on its two front wheels, according to a report on research by ETH Zurich, Intel and collaborators. The machine also managed the stunt on uneven ground, using onboard sensors rather than cameras or lidar to read the terrain.
Watch Desk analysis
What happened
The researchers trained separate behaviours in simulation, then combined them on a real John Deere 332G skid-steer loader. They used measurements and about 90 minutes of data from the machine to build a model of how its hydraulic drivetrain responds. The results were published in Science Robotics on 11 November, according to DongA Science’s report.
The system lifted the loader onto its front wheels, kept its balance as the bucket arm shifted the centre of mass, and landed back on all four wheels. It also managed to move across a 75-centimetre mound, despite being trained on flat ground. Once trained, the system ran on a standard laptop with an Intel Core i7 processor and 16GB of memory, without a high-performance GPU.
Key findings
- The loader learned without stunt demonstrations
Reinforcement learning in simulation taught the system to lift, balance, control speed and land, then combine those behaviours on the real machine. - It adapted beyond its training conditions
At a bucket-arm angle outside the training range, it succeeded in 11 of 14 trials; the researchers also report successful balancing on uneven ground. - The deployed system used modest computing hardware
The real-machine demonstrations ran on a laptop, not a large GPU-equipped system.
Why it matters
Heavy machinery is difficult to automate: hydraulic responses are complex, and repeated real-world experiments can be risky. Training in simulation and transferring the learned control to real equipment offers a way to test manoeuvres that would be awkward to learn by trial and error on a five-ton machine.
The stunt is a research demonstration, not evidence that loaders are ready to operate autonomously on building sites. But the team’s longer-term aim is more practical: combine control behaviours and add cameras, lidar and other perception tools for more complex construction tasks. Getting a machine onto a transport truck without a separate ramp is one possibility raised by the report.
Our read
The party trick is spectacular; the useful result is that the system handled a real machine, changing weight distribution and uneven terrain without leaning on a heavyweight GPU. Construction sites are not stunt arenas, of course, but this is a concrete step towards machines that can cope with their own physics rather than merely behave in a tidy simulation.
What to watch
- Whether the researchers can extend the approach from individual manoeuvres to complete construction tasks.
- How it performs across different machines and less predictable operating conditions.
- Whether later systems combine the control method with cameras or lidar, as the team plans.
Discussion spark: For autonomous construction equipment, should researchers prioritise proving reliable control on real machines first, or move sooner to integrating the cameras and lidar needed for practical site work?
Sources and evidence
- ETH Zurich and Intel AI balance 5-ton skid-steer loader on two front wheels – DongA Science (12 November 2026, 00:15 UTC)
- Source update (12 November 2026, 00:15 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.