AI two-wheel balance on a skid-steer loader leaves the off-nominal case open

ETH Zurich's Robotic Systems Lab, working with Intel, reported in Science Robotics on 11 November that reinforcement-learning control let a John Deere 332G skid-steer loader lift onto its two front wheels, balance, and land. The unresolved point is how far that balance holds once the machine leaves the conditions it trained on.

The source describes the mechanism as a learned motion model. The team measured physical properties such as length and center of mass, fed control commands into the real hydraulic drivetrain for about an hour and a half, and used the responses to build a neural network predicting motion. Reinforcement learning then ran in simulation, with varied mass, inertia and sensor error to narrow the real-to-virtual gap.

That gap showed up in the bucket arm angle. At 28 and 31 degrees the loader completed every lift trial; at 34 degrees, outside the training range, it succeeded in 11 of 14. Whether such degradation would matter in earthmoving work is a question, not a finding.

The concrete detail still to be demonstrated is generalization beyond trained configurations, which the team's stated plan to add cameras and lidar may address.