Machine Learning PapersLandmark papers
← All papers
RESEARCH PAPER / 2026-11-13

Heavy Metal Acrobatics: Learning Dynamic Control for Skid-Steer Loaders

Pierre-Jean Lorentz, Pascal Egli, Edo Jelavic, Cliff Li, Ramitha Sundar, Harikrishnan Suresh, Paul Montgomery, Lee Redden, Matthias Müller, Ozan Sener, Vladlen Koltun, Marco Hutter

ETH Zürich +3 · Science Robotics

Affiliations: ETH Zürich, Blue River Technology, Gravis Robotics AG, Intel

Abstract

Automation of heavy machinery offers the opportunity to improve productivity in mechanized industries such as construction, mining, and material handling. Yet, automation of such machines is challenged by their complex dynamics, in part because of hydraulic actuation that involves fluids, flexible elements, and nonlinear delays. A concomitant challenge is the size and weight of these machines, which can make experiments on the physical systems exceedingly dangerous. Here, we show that \ac{RL} in simulation, followed by transfer of the learned policies from simulation to reality, can successfully automate acrobatic maneuvers that push a heavy construction machine to the limits of its actuation and have only previously been accomplished by a small number of elite human operators in official world-record performances. We trained policies that perform autonomous acrobatic maneuvers on a John Deere 332G, an operational 5-tonne skid-steer loader. The policies can autonomously swing the machine up onto two wheels and, while balancing on two wheels, drive over rough terrain, trace out specified trajectories, spin around, and swing the bucket back and forth before landing it safely back onto four wheels. This work demonstrates that \ac{RL} and sim-to-real transfer enable dynamic autonomous control of heavy machinery, directly contributing to greater automation in heavy equipment sectors. Videos: https://sites.google.com/leggedrobotics.com/heavy-metal-acrobatics This is the author's version of the work. It is posted here by permission of the AAAS for personal use, not for redistribution. The definitive version was published in Science Robotics on 11 November 2026, DOI: https://doi.org/10.1126/scirobotics.aef0894

The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.

Research summary

The paper demonstrates that reinforcement learning combined with sim-to-real transfer can enable highly dynamic, acrobatic maneuvers on a full-scale 5-tonne skid-steer loader, outperforming traditional MPC baselines in robustness and velocity tracking. This work significantly advances the field of robotic control by providing a practical framework for automating complex hydraulic machinery, with clear potential for industrial applications in heavy equipment sectors.

Methodology

The paper employs a standard but effective sim-to-real reinforcement learning pipeline. The core methodological contribution lies in the practical procedure for characterizing hydraulic actuation dynamics using a neural network trained on real-world data to bridge the sim-to-real gap, rather than relying on analytical models which are often inaccurate for hydraulic systems. The authors decompose the complex acrobatic maneuver into four distinct skills (swing-up, stabilize, balance/track, soft landing), each trained separately with PPO. While the decomposition strategy is a known technique in robotics, its application to a full-scale, heavy hydraulic machine with such dynamic requirements is significant. The use of domain randomization and external disturbances during training is standard practice but effectively applied here to ensure robustness.

Results & evaluation

The experimental evaluation is strong, featuring real-world deployment on a 5-tonne John Deere 332G skid-steer loader. The paper provides quantitative comparisons against a Model Predictive Control (MPC) baseline, demonstrating that the RL policy outperforms MPC in velocity tracking and robustness to terrain disturbances. The inclusion of failure cases (e.g., bucket touching ground when arm is higher than training range) adds credibility. The visual evidence of the machine performing wheelies and spins is compelling and demonstrates the capability of the learned policies.

Reproducibility

The authors provide code and datasets via Dryad, which is a positive step. However, the specific hardware setup (John Deere 332G with specific retrofits) is not easily reproducible by the general research community. The simulation environment details (RaiSim) and hyperparameters are provided in the supplementary materials, allowing for partial reproduction of the simulation aspects. The actuator model training procedure is well-documented.

Limitations

The primary limitation is the reliance on hand-crafted switching logic between skills, which limits the scalability of the approach to more complex, long-horizon tasks. The paper acknowledges this and suggests hierarchical RL as a future direction. Additionally, the experiments are limited to a specific machine configuration and do not explore generalization to other machine types extensively, although the actuator modeling approach is general. The safety modifications (wheelie bar, bucket wheels) are necessary for testing but alter the dynamics slightly, though the authors argue they were excluded from the collision model to prevent exploitation.

Broader impact

This work has significant implications for the automation of heavy machinery in construction, mining, and agriculture. By demonstrating that RL can handle the complex, nonlinear dynamics of hydraulic systems, it opens the door for more autonomous and agile operations in these industries. The techniques developed for actuator modeling and sim-to-real transfer are applicable to other robotic systems with similar actuation challenges. The paper demonstrates that reinforcement learning combined with sim-to-real transfer can enable highly dynamic, acrobatic maneuvers on a full-scale 5-tonne skid-steer loader, outperforming traditional MPC baselines in robustness and velocity tracking. This work significantly advances the field of robotic control by providing a practical framework for automating complex hydraulic machinery, with clear potential for industrial applications in heavy equipment sectors.

Author of this paper? Send a correction.