Heavy Metal Acrobatics: Learning Dynamic Control for Skid-Steer Loaders
Authors: 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
Organizations: Robotic Systems Lab (RSL), ETH Zürich, Zürich 8092, Switzerland · Blue River Technology (BRT), Santa Clara, CA 95054, USA · Gravis Robotics AG, Opfikon 8152, Switzerland · Intelligent Systems Lab, Intel, Jackson, WY 83001, USA
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
Figures & tables
Figure 1: Acrobatic maneuvers trained with RL in simulation and deployed on a 5-tonne SSL . (A) Swing-up from four onto two wheels. (B) Driving over rough terrain on two wheels. (C) Stable balancing while moving the arm. (D) Spinning in place on the front wheels. (E) Soft landing after balancing on two wheels.
Figure 2: Hardware description. (A) John Deere 332G SSL . For safety, a wheelie bar with a rolling tube was mounted at the back (1) and wheels were mounted on the bucket (2). Two IMU were added: one on the chassis (not visible) and another one on the arm to monitor its angle (3). Inside the machine, two internet devices (4) and (7) ensured wireless connectivity for the laptop (6), which ran the controller and interfaced the machine’s CAN bus. A physical switch (5) allowed us to enable autonomous mode only when needed. (B) Different positions of the arm. (I) is the upper joint limit ( 34∘ ), and (II) is the lowest position possible to maintain balance without touching the ground ( 28∘ ). (C) Underside view of the wheelie bar (I) and side view during swing-up (II). Note that the wheelie bar does not change the machine’s ground clearance and remains unloaded during swing-up as it is bent upward.
Figure 3: Hardware deployment. (A) Image sequence of the swing-up maneuver. (B) Swing-up in simulation and reality. First, the swing-up controller is executed. When the machine’s pitch reaches −30\mathrm{\SIUnitSymbolDegree} (dashed lines), the stabilize-and-balance controller is activated. The roman numerals refer to the corresponding images of (A). (C) Soft-landing sequence in simulation and on the real SSL compared with zero command. In the beginning, the balancing controller is active. The landing controller is triggered at $t=1\penalty\ $\mathrm{s} and is active after that. When the pitch, pitch rate, and wheel speed are close to zero, the commands are set to zero. If the soft-landing controller is inactive and the action is set to zero immediately after balancing, the SSL falls onto four wheels. However, the front wheels rebound, causing the machine to rock back and forth. (D) Image sequence of the soft-landing maneuver with the roman numerals corresponding to the time stamps in (C). (E) Balancing in place with RL in simulation and on the real machine, with the MPC baseline. (F) Velocity tracking with RL and MPC . The target wheel velocity is increased/decreased gradually up to the maximal/minimal velocity that each controller can successfully track without touching the ground with the bucket or the rear underside. Positive target velocity tracking starts with the swing-up maneuver. Negative target velocity tracking starts with the transient behavior of switching from a positive target to a zero-velocity target.
Figure 4: Robustness analysis. For all panels, the corresponding snapshot times are indicated in the plot with roman numerals. (A) Robustness to arm movements of the RL controller and the MPC baseline. The arm is moved around its nominal position. Between (I) and (V), the arm moves downward; between (V) and (IX), the arm moves upward. (B) Robustness to spinning at different wheel speeds of the RL policy and the MPC . Spinning is achieved by overlaying a machine rotation command on the policies’ actions. (C) Traversing rough terrain. The SSL successfully navigates the 75\penalty\mathrm{cm}$$ bumps, driving forward and backward. The image sequence shows the SSL driving backward over the bumps. The green curve shows the MPC not being able to traverse the first bump, even when increasing the target velocity.
Figure 5: Method overview. (A) The controllers were first trained in simulation and then deployed on real hardware. The simulation incorporated an actuator model trained to predict wheel speeds on the basis of real-world data and a rigid-body model whose properties were randomized during training around the measured values. (B) Wheel speed for the same joystick input, rotating freely in the air (machine is lifted with a car jack) and when placed on the ground (thus moving the entire SSL ). (C) Comparison of the wheel speed in simulation and on the real machine for different types of signals: ramps, sinusoidal, and commands from the controller while balancing the machine. (D) Hand-measured collision geometry of the rigid-body model (highlighted in red). Generic mesh for visualization only in gray. (E) Visualization of the wheel and machine centers of mass (red dots) and the inertia approximation with a uniform mass distribution over the red cuboid for the cabin and four cylinders for the wheels. Note that the machine’s center of mass changes depending on the arm configuration.
Figure 6: RL. (A) Initial state distribution at episode start (sampled on 1000 environments). (B) Visualization of different SSL pitch angles. (C) Rewards as a function of the pitch angle for swing-up and soft-landing controllers for different values of c0,1 . (D) Swing-up learning: Average reward per time step and curriculum progression on c0 to increasingly penalize touching the ground with the rear underside. (E) Distribution of the pitch angle at the end of an episode. All 1000 trials achieved a pitch angle smaller than −30∘ , which is required for successful swing-up. (F) Stabilize and balance learning: The termination probability drops from 1 to 0, corresponding to the transition from failing all of the time to successful balancing without touching the ground. The rewards for the command and the command derivative increase over the course of training as the controller learns how to perform the task using small and smooth actions. (There is no curriculum for this task.) (G) Soft landing learning: The probability of terminating by landing on the bucket or tipping over backward drops to 0. The average reward per time step increases. (There is no curriculum for this task.) (H) Balance and track velocities learning: The probability of falling when balancing and tracking velocities drops from 1 to 0. The tracking reward decreases because of the curriculum on c3 .
Figure 7: Mathematical description of the SSL used for the MPC . Each front wheel is modeled as a cylindrical body with radius R , mass mwheel , and inertia Θwheel . The rest of the machine is lumped into a body of mass mbody and inertia Θbody .
Figure 8: SMB ( SMB ). (A) The SMB is a skid-steer vehicle for heavy-duty use in challenging conditions. It has been retrofitted with a 6- DoF IMU , a PC to run the policy, read the IMU signals and communicate with the motor controller, an emergency stop for safety, a roll cage for protection of the components, and wheel encoders on the front wheels. (Bi) The operator is navigating the SMB accurately along the path by providing high-level commands consisting of linear and angular velocity. The policy is robust to uneven ground and driving down small steps. (Bii) The policy can reject external pushes, which can cause the SMB to travel a couple of meters. (Biii) the policy is robust to non-flat ground and can successfully balance and follow tracking commands. (C) Control command and velocity tracking performance in simulation and on the real system while balancing, given the same command sequence. (D) The hybrid simulation of the SMB consists of an actuator model trained to predict torques based on real-world data, and the rigid-body model whose properties are estimated and randomized around the measured values during training. (E) Actuator model performance evaluation on a validation dataset in simulation compared to measurements on the real machine.
Humanoid robots could transport payloads substantially heavier than themselves by pulling passive wheeled vehicles instead of carrying the load. This capability, however, creates a coupled locomotion problem: the robot must maintain persistent upper-body contact while adapting to unknown, configuration-dependent forces arising from the payload, vehicle, and terrain. We present a whole-body control framework for humanoid rickshaw pulling that tracks commanded vehicle motion while preserving balance and stable grasps under uncertain load dynamics. During training, a privileged teacher exploits vehicle states, interaction forces, and load properties. Its actions and latent are distilled into a history-conditioned student that implicitly infers coupled dynamics from proprioceptive responses, followed by reinforcement-learning fine-tuning. Comparisons with No History and Only History baselines show that the resulting policy achieves accurate vehicle tracking while reducing vehicle oscillation, torso tilt, and actuation cost. Behavioral analysis shows that Unitree G1 propels the rickshaw and generates gait-synchronized whole-body reactions that stabilize its lateral and roll motions. Moreover, pulling redistributes joint effort and yields a lower robot-normalized cost-of-transport proxy than unloaded walking over most tested load--speed conditions. On hardware, a single policy performs starting, sustained pulling, turning, and stopping with both rigid payloads and human passengers, handling a loaded rickshaw mass of up to 115~kg without load-specific retuning. These results demonstrate robust heavy-load transportation through coordinated and persistent humanoid--vehicle interaction.
Precise, high-speed control remains challenging for robots with complex actuation dynamics. Learning directly on hardware is further constrained by the cost of real-world interaction. We present an online model-based reinforcement learning framework that learns a probabilistic dynamics ensemble model from scratch for sampling-based model predictive control. A precision-gated contouring objective conditions the progress reward on path accuracy, prioritizing precision over speed. In a data-driven excavator simulator, the framework achieves higher sample efficiency than the evaluated model-based reinforcement learning baselines. We validate the framework by learning directly on an 11.5-ton Menzi Muck M445 hydraulic excavator, without demonstrations or simulation pretraining. After 20 minutes of interaction, the controller reaches tracking accuracy comparable to prior learned controllers trained on 100-150 minutes of data. After 40 minutes, it sustains sub-centimeter mean path error at high operating speeds.
Claudio Canales, Fang Nan, Marco Hutter +1
Robotic Systems Lab, ETH Zürich, 8092 Zürich, Switzerland · Advanced Mining Technology Center (AMTC) and the Department of Electrical Engineering, Universidad de Chile, Santiago, Chile
Autonomous excavator control is challenged by coupled kinematics, actuation lag, and uncertainty. We propose imitation learning and adaptive Cartesian tracking (IL-ACT), a novel motion control framework for a 30-ton-class excavator. An anchored, 14-input imitation policy pretrained on operator demonstrations generates nominal joint rates; adaptive Cartesian feedback and gated gain/bias estimation correct these commands before a stopping-distance governor constrains joint-reference generation. Simscape evaluation covers 100 sequential goals and spiral, figure-eight, and rounded-raster tracking, including 88 additional runs across three training seeds, two initializations, and speeds, under hydraulic response and sensing conditions. Compared with Teacher+ACT, IL-ACT completes all goals with shorter duration and lower terminal errors under both response conditions. Telemetry-initialized IL-ACT lowers RMSE in all 24 figure-eight and rounded-raster seed comparisons and lowers additional-load spiral mean RMSE by approximately 29%. Original spiral RMSE also improves over IL-only and PID. Under a shared sensor-noise realization, telemetry-initialized IL-ACT achieves 27.67% lower mean RMSE than Teacher+ACT; enabling estimation reduces mean RMSE by 22.44% relative to the frozen estimator. Pretrained-weight effects remain mixed, and the original teacher comparison exhibits a spiral RMSE--maximum-error tradeoff. Analysis establishes bounded adaptive states and Cartesian feedback, with reference admissibility conditional on governor feasibility.
Mehdi Heydari Shahna, Seihun Kim, Soyi Jung +3
Faculty of Engineering and Natural Sciences, Tampere University, Tampere, Finland · Department of Defense Convergence Technology, Korea University, Seoul, Republic of Korea · Department of Electrical and Computer Engineering, Korea University, Seoul, Republic of Korea +2