Mechanical Control Systems Overview

• Mechanical Control Systems are defined by second-order equations on configuration manifolds, preserving true configurations, velocities, and force interpretation.
• Modern approaches employ geometric tracking, Lie-group formulations, and structure-preserving linearization to ensure robust and accurate control in both fully actuated and underactuated settings.
• Applications span robotics, aerospace, and distributed coordination, leveraging port-Hamiltonian dynamics and energy shaping to achieve high-performance, structure-consistent control.

Mechanical Control Systems (MCS) are control systems whose plant dynamics arise from mechanical laws and are naturally expressed as second-order equations on a configuration manifold rather than as arbitrary first-order state equations in Euclidean coordinates. In contemporary formulations, the state evolves on the tangent bundle $TQ$ or $TG$, inertia is encoded by a Riemannian metric or affine connection, forces appear as vector fields or port variables, and control inputs retain a force/torque interpretation under admissible transformations (Nowicki et al., 2022, Maithripala et al., 2016, Chan-Zheng et al., 2021). The area encompasses fully actuated and underactuated systems, Lie-group models for rigid-body motion, port-Hamiltonian energy-based descriptions, geometric tracking on manifolds, structure-preserving linearization, adaptive and finite-time control, distributed coordination, and variational optimal control (Espindola et al., 2023, Ferguson, 2023, Javanmardi et al., 2022).

1. Intrinsic formulations and system classes

A canonical geometric model for a single-input mechanical control system is the $4$-tuple $(Q,\nabla,g,e)$, with trajectories $x(t)\in Q$ satisfying

$\nabla_{\dot{x}(t)}\dot{x}(t)=e(x(t))+g(x(t))u,$

where $\nabla$ is a symmetric affine connection, $e$ is the uncontrolled force field, and $g$ is the controlled force field (Nowicki et al., 2022). In local coordinates this becomes a first-order system on $TQ$,

$TG$0

so the state is intrinsically a position–velocity pair. This formulation includes standard Lagrangian systems without dissipation, but it is explicitly more general because $TG$1 need not come from a metric and $TG$2 need not be a gradient field (Nowicki et al., 2022).

A second major language is the port-Hamiltonian description. For generalized coordinates $TG$3, momenta $TG$4, inertia matrix $TG$5, and Hamiltonian

$TG$6

the plant dynamics are written through a skew interconnection, damping, and input port, with passive output typically given by $TG$7 or $TG$8 (Ferguson, 2023, Chan-Zheng et al., 2021). This representation is mechanically equivalent to Euler–Lagrange dynamics under the Legendre transform $TG$9, but it makes stored energy, dissipation, and interconnection structure explicit.

For systems on manifolds with symmetry, Lie groups are central. A configuration $4$0 has generalized velocity in $4$1, and many rigid-body models are more naturally written using left or right trivializations, such as $4$2 or $4$3 (Maithripala et al., 2016). In that setting, the intrinsic mechanical equation is

$4$4

or, after trivialization, $4$5, where $4$6 is the inertia map induced by the kinetic-energy metric (Maithripala et al., 2016). Affine-connection formulations of extremum seeking use the same geometric acceleration concept: $4$7 with $4$8 a properly embedded submanifold of Euclidean space (Suttner, 2021).

Taken together, these formulations show that MCS are defined less by a particular coordinate form than by preservation of mechanical meaning: true configurations, true velocities, inertial geometry, and force-like actuation.

2. Geometric tracking and manifold-compatible feedback

A persistent theme in MCS is that coordinate-based control can destroy global geometric meaning. For fully actuated mechanical systems on Lie groups, geometric PID control treats the dynamics as a generalized double integrator on a manifold and replaces Euclidean error coordinates by invariant group errors $4$9 or $(Q,\nabla,g,e)$0, with corresponding velocity errors $(Q,\nabla,g,e)$1 (Maithripala et al., 2016). The proportional term is derived from a polar Morse function $(Q,\nabla,g,e)$2, the derivative term uses $(Q,\nabla,g,e)$3, and the integral action is introduced through the covariant law

$(Q,\nabla,g,e)$4

The resulting PID law preserves the mechanical and geometric structure of the plant and yields robust almost-global, locally exponential tracking under the stated gain conditions (Maithripala et al., 2016).

The same geometric concern appears in sliding-mode design. For fully actuated simple mechanical systems whose configuration manifold is a Lie group, a geometric sliding-mode controller constructs a Lie-group structure on the tangent bundle $(Q,\nabla,g,e)$5, then defines a sliding subgroup compatible with the true state-space geometry (Espindola et al., 2023). This is not a cosmetic reformulation of Euclidean sliding mode using attitude coordinates. The paper’s central point is that if a sliding set is not properly embedded in $(Q,\nabla,g,e)$6, the sliding motion may fail to exist as a geometric object. The proposed controller ensures that error trajectories reach the sliding subgroup globally exponentially during the reaching phase, and that once on the subgroup the reduced error dynamics yields almost-global asymptotic tracking and local exponential tracking (Espindola et al., 2023).

This directly addresses a common misconception: Euclidean control ideas do not transfer automatically to manifolds. In MCS on $(Q,\nabla,g,e)$7, $(Q,\nabla,g,e)$8, or other nonlinear configuration spaces, singularities, ambiguities, or topological obstructions are not secondary implementation issues; they are structural features of the control problem (Espindola et al., 2023, Maithripala et al., 2016).

3. Structure-preserving linearization and decoupling

Classical feedback linearization asks whether a nonlinear state-space model can be mapped to a linear controllable system. Mechanical feedback linearization asks a stricter question: whether a mechanical system can be transformed into a linear mechanical system by transformations that preserve configuration coordinates, true velocities, and force interpretation (Nowicki et al., 2022). The admissible coordinate changes are tangent lifts

$(Q,\nabla,g,e)$9

and the admissible feedbacks have the mechanical form

$x(t)\in Q$0

For scalar-input systems, the theorem of Nowicki–Respondek gives necessary and sufficient conditions for local MF-linearizability, combining the classical configuration-space accessibility conditions with additional compatibility conditions involving $x(t)\in Q$1 and $x(t)\in Q$2 (Nowicki et al., 2022). A notable feature is verifiability: the test uses differentiations and algebraic operations only.

An allied development concerns simultaneous input-output linearization and decoupling. For square mechanical systems with outputs $x(t)\in Q$3, the Mechanical Input-Output Linearization and Decoupling problem requires that both the transformations and the resulting normal form remain mechanical (Nowicki et al., 2023). The target is not the generic Byrnes–Isidori normal form but a decoupled mechanical chain of configuration/velocity pairs. Solvability holds if and only if two conditions are satisfied: a well-defined vector relative half-degree $x(t)\in Q$4, and the covariant constancy condition

$x(t)\in Q$5

The first condition resembles classical decoupling; the second is specific to MCS and forces the output-generated coordinates to be affine with respect to the connection (Nowicki et al., 2023).

These results show that the mechanically admissible class is strictly smaller than the class of general feedback-linearizable systems. A system may be classically decouplable yet fail mechanically because the usual normal form mixes positions and velocities in a way that destroys second-order mechanical structure (Nowicki et al., 2023). This suggests that in MCS, linearization is best viewed as a structure-preserving equivalence problem rather than a purely input–output one.

4. Energy shaping, passivity, and damping-aware design

Energy-based control is one of the most developed strands of MCS. In underactuated port-Hamiltonian systems, total energy shaping has traditionally been limited by the solvability of IDA-PBC matching conditions. A control-by-interconnection interpretation recasts the problem by introducing a dynamic passive controller whose states are locked to the plant by Casimirs, so that the controller stores an added kinetic energy and added potential energy (Ferguson, 2023). The closed-loop inverse mass is

$x(t)\in Q$6

and the matching conditions are rewritten in terms of the added inverse mass $x(t)\in Q$7 and $x(t)\in Q$8, rather than the total closed-loop energy. For degree-one underactuation and mass matrices depending on a single coordinate, the kinetic matching conditions reduce from PDEs to ODEs that can be evaluated numerically; the approach is illustrated on the cart-pole and acrobot (Ferguson, 2023).

Passivity-based stabilization also appears in a more implementation-oriented form. For nonlinear mechanical systems in port-Hamiltonian form, linear viscous damping can be identified from data via an energy-based regression derived from the momentum equations, and the resulting estimate can then be used to tune PI-PBC or modified PI-PBC controllers to reduce transient oscillations (Chan-Zheng et al., 2021). The method focuses on the linear component of the natural damping matrix $x(t)\in Q$9, because the tuning rules depend on that component. A planar manipulator experiment is used for validation (Chan-Zheng et al., 2021).

At the network level, timed IDA-PBC is extended to large-scale formations of fully actuated mechanical agents modeled as port-Hamiltonian systems with constant mass matrix $\nabla_{\dot{x}(t)}\dot{x}(t)=e(x(t))+g(x(t))u,$0. Each follower uses only local neighbor information and prescribed offsets $\nabla_{\dot{x}(t)}\dot{x}(t)=e(x(t))+g(x(t))u,$1, while a PDE-based approximation of the network establishes scalable asymptotic stability under homogeneous gains (Javanmardi et al., 2022). Here energy shaping and damping injection remain local agent-level design tools, but the analysis shifts to a distributed-parameter viewpoint.

5. Robust, adaptive, and learning-oriented feedback

Robust regulation in MCS often turns on the tension between disturbance rejection and chattering. A discontinuous integral controller for the perturbed double integrator

$\nabla_{\dot{x}(t)}\dot{x}(t)=e(x(t))+g(x(t))u,$2

combines a homogeneous non-Lipschitz state feedback with a discontinuous integral state $\nabla_{\dot{x}(t)}\dot{x}(t)=e(x(t))+g(x(t))u,$3 and, when needed, a continuous finite-time velocity observer (Moreno, 2015). The plant input remains continuous even though $\nabla_{\dot{x}(t)}\dot{x}(t)=e(x(t))+g(x(t))u,$4 is discontinuous, and for Lipschitz perturbations the closed-loop equilibrium is globally finite-time stable. The controller can be interpreted as a perturbation estimator because $\nabla_{\dot{x}(t)}\dot{x}(t)=e(x(t))+g(x(t))u,$5 after finite time (Moreno, 2015).

For underactuated Lagrangian systems, collocated adaptive control extends the Slotine–Li framework by introducing a fictitious input $\nabla_{\dot{x}(t)}\dot{x}(t)=e(x(t))+g(x(t))u,$6 for the unactuated coordinates, thereby recovering linear parameterization at the level of the full system dynamics (Romano et al., 2014). The resulting controller requires no acceleration measurements, proves local stability, and gives convergence of the collocated tracking error to zero under an additional boundedness assumption on the noncollocated velocities (Romano et al., 2014). This is significant because eliminating unactuated accelerations from the reduced collocated dynamics destroys the linear-in-parameters structure on which standard adaptive robot control depends.

Extremum seeking provides a different route when the objective is not reference tracking but stabilization around minima of an output function $\nabla_{\dot{x}(t)}\dot{x}(t)=e(x(t))+g(x(t))u,$7 using only real-time measurements of $\nabla_{\dot{x}(t)}\dot{x}(t)=e(x(t))+g(x(t))u,$8. For affine-connection mechanical systems, high-frequency periodic perturbations and a high-pass filter yield an averaged system whose descent directions are generated by symmetric products rather than Lie brackets (Suttner, 2021). Under suitable assumptions, minima of $\nabla_{\dot{x}(t)}\dot{x}(t)=e(x(t))+g(x(t))u,$9 are asymptotically stable for the averaged system and practically asymptotically stable for the oscillatory closed-loop system. This is a specifically mechanical version of geometric extremum seeking, because the second-order structure forces averaging to occur on the tangent bundle.

6. Optimal control, variational discretization, and coordinated motion

Optimal control in MCS is closely tied to variational structure. For underactuated systems with symmetry on a trivial principal bundle $\nabla$0, optimal control problems can be rewritten as higher-order variational problems with higher-order constraints, and then discretized into variational integrators that preserve symplecticity, momentum preservation, and good behavior of the energy (Colombo et al., 2012). The same geometric viewpoint underlies discrete optimal control of interconnected underactuated systems on product Lie groups $\nabla$1, where a discrete Lagrangian and forced discrete Euler–Lagrange equations produce a structure-preserving discrete dynamics model, and the resulting boundary-value problem is solved by multiple shooting (Nair et al., 2018).

A continuous optimal-control development goes further by interpreting manipulator trajectory generation on a Riemannian configuration manifold. With dynamics

$\nabla$2

the Pontryagin system reveals three geometric effects on optimal trajectories: curvature effects of the inertia manifold, curvature effects of the potential field, and shortening effects from resistive force (Choi et al., 2024). The pure inertial limit reduces to the Riemannian spline equation

$\nabla$3

while gravity enters through the Riemannian Hessian of the potential and drag contributes a path-shortening tendency through an induced drag metric (Choi et al., 2024). This suggests that mechanically optimal motion is shaped not only by the target path but by morphology, gravity, and dissipation.

Distributed coordination is a further extension of the same structure-conscious perspective. Leader–follower formation control for networked mechanical systems uses local pH/tIDA-PBC design, then interprets the mesh-like communication graph as a spatial discretization of a PDE, yielding scalable asymptotic stability independent of network size under the stated assumptions (Javanmardi et al., 2022). Here MCS expands from single rigid bodies or manipulators to large coordinated populations.

7. Applications, experiments, and recurring limitations

The application range documented in the literature is broad. Geometric tracking and sliding-mode results are demonstrated on attitude control over $\nabla$4 and $\nabla$5 (Espindola et al., 2023). Geometric PID is illustrated on a multi-rotor aerial vehicle, a hoop rolling on an inclined plane, a sphere rolling on an inclined plane, and a spherical pendulum (Maithripala et al., 2016). Energy-shaping methods are tested on cart-pole and acrobot benchmarks (Ferguson, 2023). Distributed formation control is demonstrated on six spacecraft in a $\nabla$6-dimensional mesh-like graph (Javanmardi et al., 2022). Discrete variational optimal control is illustrated on the ball-and-beam and inverted pendulum on a cart (Nair et al., 2018). Damping identification is validated on a planar manipulator (Chan-Zheng et al., 2021). A tabletop active pendulum with switchable length, driven by solenoids, serves as a mechatronic demonstration of sway attenuation and swing pumping through event-based length modulation (Stein et al., 2023).

The same body of work also defines the field’s practical boundaries. Many results assume full actuation, constant or accurately known mass matrices, ideal communication, smooth outputs that depend only on configuration, or high-frequency perturbations (Javanmardi et al., 2022, Nowicki et al., 2023, Suttner, 2021). Several linearization and decoupling theorems are local in configuration, even if global in velocity (Nowicki et al., 2023). Adaptive and robust schemes may require bounded noncollocated velocities or perturbations with bounded derivative (Romano et al., 2014, Moreno, 2015). The distributed formation result explicitly does not cover delays, packet drops, or configuration-dependent inertia (Javanmardi et al., 2022). The active pendulum experiment reports close agreement between simulation and experiment, but also reveals actuator dead time, impact-induced vibration, and passive extension limits as dominant nonidealities (Stein et al., 2023).

A plausible synthesis is that modern MCS research is organized around one overarching principle: control laws, transformations, discretizations, and performance analyses are strongest when they preserve the plant’s mechanical structure rather than suppressing it. In that sense, the field has moved from treating mechanics as a nuisance nonlinearity toward treating it as the primary source of coordinates, invariants, admissible feedbacks, and optimality conditions (Nowicki et al., 2022, Nowicki et al., 2023, Choi et al., 2024).