Hydraulic robots could soon handle heavy machinery with the steadiness of far lighter electric systems, according to a new control strategy designed to make six-degree-of-freedom machines more accurate, safer during contact and less dependent on costly force sensors. The approach, developed by researchers at Xidian University in China, combines virtual decomposition control with an indirect method for estimating forces at a robot’s end effector. In simulations, the system achieved millimetre-level positioning accuracy and substantially reduced peak contact forces compared with a conventional controller. The results point to a possible route for improving robots used in construction, aerospace manufacturing, underwater operations, rehabilitation and other environments where hydraulic power is essential but precision is difficult to maintain. The study was published in the International Journal of Intelligent Robotics and Applications on 26 August 2026.
Hydraulic actuation remains attractive for machines that must move heavy loads because pressurised fluid can deliver very high force from relatively compact cylinders and motors. That high power-to-weight ratio is valuable in excavators, cranes, aircraft-service equipment and large manipulators, but it comes with a difficult control problem. Unlike an ideal electric motor, a hydraulic actuator is affected by compressible fluid, valve flow nonlinearities, pressure changes, seal friction and leakage. The payload can vary from one task to the next, while external forces may arrive suddenly when a tool touches an object. These effects interact with the robot’s rigid-body dynamics, meaning that motion in one joint can influence forces and errors throughout the entire six-axis mechanism. A controller that works well in free space may therefore overshoot, vibrate or press too hard when the robot encounters resistance.
The researchers addressed this complexity by building a coupled model that represents both sides of the machine: the mechanical structure and the hydraulic actuators that drive it. The rigid-body portion describes joint positions, velocities, accelerations, link masses and the forces transmitted through the robot. The actuator model captures fluid dynamics and hydraulic friction, allowing the controller to account for how pressure and flow generate cylinder forces rather than treating each actuator as a simple ideal motor. This distinction is technically important. In a hydraulic cylinder, the commanded valve opening does not translate instantly into a precise mechanical force. Fluid compressibility, changing chamber pressures and friction can introduce delays and disturbances, so accurate control requires the robot’s mechanical and fluid behaviour to be considered together.
At the centre of the proposed system is virtual decomposition control, or VDC. Rather than controlling the entire robot as one enormous nonlinear system, VDC conceptually divides it into interconnected subsystems. Individual links, joints, actuators and transmission elements can be analysed and controlled locally, while the forces and velocities exchanged at their interfaces are coordinated to preserve the behaviour of the complete robot. The decomposition is “virtual” because the physical robot is not cut into separate machines; instead, its mathematical description is partitioned into manageable units. Each subsystem receives a control objective, and interface variables maintain consistency between neighbouring units. This structure can reduce the burden of deriving and tuning a single global controller, especially for multi-joint hydraulic robots whose equations contain strong coupling and uncertain parameters.
The resulting controller was designed to track trajectories in both joint space and Cartesian space. Joint-space tracking asks each actuator to follow a specified angle or displacement, whereas Cartesian tracking focuses on the position of the tool in three-dimensional space. The latter is often more useful in practical work: a welding head, gripper or drilling tool must reach a target location and orientation even if the individual joints take different paths. The controller generates the coordinated actuator commands needed to reduce tracking errors while compensating for nonlinear dynamics. In the reported simulations, the robot was tested under trapezoidal velocity planning, a common trajectory profile that accelerates and decelerates in controlled stages rather than changing speed abruptly. Along the x-axis, the maximum Cartesian error was 1.156 millimetres and the root-mean-square error was 0.525 millimetres.
Those figures are especially notable because the tests were not limited to an unloaded robot moving smoothly through empty space. The study also examined how the controller could estimate generalized forces without a force sensor mounted at the end effector. A generalized force is a compact representation of the forces and torques acting on a robot’s motion coordinates. In Cartesian operation, it can include the three components of linear force and the three components of torque; in other formulations, it can describe the equivalent forces associated with joint motion. Direct measurement normally requires a six-axis force-torque sensor near the tool, but such sensors can be expensive, fragile and vulnerable to overload in heavy-duty environments. They can also add mass at the end of a long manipulator, potentially reducing payload capacity and changing the dynamics that the controller must manage.
The sensorless estimator instead infers external forces from quantities already available within the robot’s control system, such as joint motion, actuator pressures and the mathematical model of the mechanism. In principle, the controller compares the forces predicted by the robot’s internal dynamics with the forces implied by measured actuator behaviour. The difference can indicate an external load or contact event. This is challenging because the same mismatch may also be caused by friction uncertainty, leakage, modelling errors or disturbances in the hydraulic circuit. The VDC framework gives the estimator a subsystem-level structure in which internal interactions can be accounted for before the remaining generalized force is attributed to the environment. The researchers validated the method under four representative scenarios: fixed-point control with trapezoidal and sinusoidal loads, and dynamic tracking with trapezoidal and sinusoidal contact forces.
Contact control is a critical test for any heavy hydraulic robot. A position controller that is excellent in free space can become unsafe when it continues trying to reach a target after the tool has touched a rigid surface. The result can be a sharp force spike, damage to the tool or workpiece, and destabilizing vibrations that travel through the manipulator. To address this problem, the researchers added variable impedance control. Impedance control changes the apparent mechanical relationship between position and force, making the robot behave as though it has adjustable stiffness, damping or inertia. A high virtual stiffness helps the robot maintain position, while lower stiffness allows it to yield when contact forces rise. Variable impedance makes those properties change according to operating conditions rather than keeping them fixed. The study reports that this strategy reduced the peak contact force along the x-axis by 1,156.254 newtons compared with a baseline that used no impedance control.
The need for variable rather than constant impedance emerged from a trade-off in the simulations. A robot tuned to be compliant during contact may absorb impacts effectively, but the same softness can degrade its free-space tracking performance. Conversely, a stiff controller can follow trajectories precisely but transmit large forces during collisions or constrained motion. The proposed approach seeks to switch or adjust the impedance so that the robot remains accurate while moving freely and more yielding when interacting with the environment. Because the force is estimated indirectly, the system can use the inferred contact condition to modify its behaviour without installing a dedicated end-effector sensor. The researchers’ results suggest that combining force estimation and variable impedance within the VDC architecture can improve both motion and interaction performance, although the reported evidence comes from simulation rather than a physical robot.
That limitation matters as the method moves toward real-world evaluation. Hydraulic systems contain unmodelled effects that can be difficult to reproduce numerically, including temperature-dependent fluid viscosity, valve dead zones, structural flexibility, pressure shocks and wear in seals. Sensor noise and delays can also corrupt force estimates, while abrupt changes in payload may expose weaknesses in the dynamic model. The authors report that no datasets were generated or analysed during the study, and they state that the control code will be provided upon request. Their work therefore demonstrates a control design and simulation-based validation, not yet a field-tested system. Experiments on a six-degree-of-freedom hydraulic platform will be needed to determine whether the millimetre-scale tracking and force reductions persist under changing loads, imperfect calibration and unpredictable contact.
Even so, the research tackles a bottleneck that has limited the autonomy of powerful hydraulic machines: they can generate enormous force, but controlling that force with finesse is difficult. A successful sensorless strategy could reduce hardware complexity while helping heavy robots distinguish between following a path and safely responding to an obstacle. Potential applications include automated handling of large aerospace components, construction and forestry machinery, remote or underwater manipulation, and assistive devices that must regulate contact with the human body. The next challenge will be proving that the estimator remains stable when the robot’s model is wrong and the environment is uncertain. If that challenge can be met, virtual decomposition control could provide a scalable framework for coordinating the mechanical and hydraulic layers of large robots, bringing industrial machines closer to the responsive behaviour expected from modern collaborative systems.
