Adaptive-Neuro Geometric Control

• Adaptive-neuro geometric control is a hybrid design combining geometric control on nonlinear manifolds with adaptive and neural disturbance compensation, ensuring robust performance in complex systems.
• It employs coordinate-free error definitions on SO(3), SE(3), or T*SE(3) to avoid singularities and support large-angle maneuvers and aggressive trajectory tracking.
• The approach integrates online neural adaptation, Lyapunov-based stability analysis, and disturbance rejection strategies, enabling practical implementations in quadrotor, Hamiltonian, and multi-agent systems.

Adaptive-neuro geometric control is a class of control designs in which geometric control on nonlinear configuration manifolds is combined with adaptive disturbance compensation and neural or learned disturbance models. In the literature, the term appears explicitly for a centralized multi-quadrotor transportation system whose geometric controller is augmented with multiple neural networks and adaptive laws, while closely related work on quadrotors, rigid-body attitude dynamics, and Hamiltonian systems develops the same synthesis under labels such as geometric adaptive control with neural networks or geometric adaptive control with learned disturbance features (Gao et al., 2 Sep 2025). Across these formulations, the recurring structure is a coordinate-free tracking law on $SO(3)$, $SE(3)$, $T^*SE(3)$, or related manifolds; an additive disturbance-cancellation term estimated online; and a Lyapunov argument that includes both geometric tracking errors and estimator or neural-weight errors (Bisheban et al., 2019).

1. Genealogy and conceptual scope

The development of adaptive-neuro geometric control is not uniform. Earlier geometric adaptive controllers on manifolds addressed matched or linearly parameterized uncertainties without neural networks. Representative examples include rigid-body payload transportation by multiple quadrotors on $$\mathbb{R}^3 \times SO(3)\times (\mathbb{S}^2 \times SO(3))^n$$, where the coupled dynamics of payload, links, and quadrotors are incorporated directly into a coordinate-free adaptive controller, and rigid-body attitude regulation on $SO(3)$ with state inequality constraints and unknown disturbances (Lee, 2015, Kulumani et al., 2016). In these papers, the adaptive term is parametric and Lyapunov-designed, but the geometric backbone is already fully developed.

A later strand augments geometric tracking laws with multilayer neural networks adjusted online to reject unknown, unstructured disturbances. For quadrotors in wind fields, the disturbances are introduced as unknown force and moment terms $\Delta_1,\Delta_2$, and neural networks are inserted directly into the translational and rotational control channels on $SE(3)$ (Bisheban et al., 2018, Bisheban et al., 2019). Another strand separates learning and adaptation: disturbance features are identified offline by a Hamiltonian neural ODE, while online adaptation estimates only the coefficients of the learned feature map in an $SE(3)$ port-Hamiltonian controller (Duong et al., 2021).

This suggests two broad interpretations of the term. In a narrow sense, it denotes manifold-based controllers whose adaptive and neural components are integrated into a coordinate-free Lyapunov analysis. In a broader sense, it also includes structured mechanical or task-driven controllers in which geometry enters through constraints, planning, or state-space structure, even when the controller is not formulated intrinsically on $SO(3)$ or $SE(3)$.

2. Geometric state spaces and intrinsic error systems

The geometric component is defined by the configuration manifold and by error variables that respect that manifold. For quadrotors, the configuration is $SE(3)$0, with $SE(3)$1, $SE(3)$2, translational velocity $SE(3)$3, and body angular velocity $SE(3)$4. For attitude-only rigid-body problems, the state evolves directly on $SE(3)$5. For Hamiltonian rigid-body systems, the state is $SE(3)$6, with generalized coordinates collecting position and orientation and generalized momentum defined by the mass matrix (Bisheban et al., 2018, Kulumani et al., 2016, Duong et al., 2021).

A defining feature is the use of manifold-consistent attitude errors. For example, the quadrotor papers define

$SE(3)$7

together with the configuration error function

$SE(3)$8

and translational errors $SE(3)$9, $T^*SE(3)$0 (Bisheban et al., 2018, Bisheban et al., 2019). The attitude-constrained $T^*SE(3)$1 controller similarly defines a trace-form attractive term $T^*SE(3)$2, a logarithmic barrier $T^*SE(3)$3, and the combined error vector $T^*SE(3)$4, so that stability and constraint avoidance are encoded in one scalar potential $T^*SE(3)$5 (Kulumani et al., 2016). In the Hamiltonian $T^*SE(3)$6 framework, the geometric tracking errors use a body-frame position error $T^*SE(3)$7, an intrinsic attitude error $T^*SE(3)$8, and momentum or velocity errors expressed in body coordinates (Duong et al., 2021).

The reason for working on these manifolds is explicit in the literature. Controllers developed directly on $T^*SE(3)$9 or $$\mathbb{R}^3 \times SO(3)\times (\mathbb{S}^2 \times SO(3))^n$$0 avoid the singularities of Euler angles, the double covering of quaternions, and unwinding phenomena, while remaining valid for large-angle maneuvers and aggressive trajectory tracking (Kulumani et al., 2016, Bisheban et al., 2019). A notable refinement is that some controllers still use Euler angles internally as neural-network inputs, but only for the approximator; the control law and the stability proof remain geometric on $$\mathbb{R}^3 \times SO(3)\times (\mathbb{S}^2 \times SO(3))^n$$1 (Bisheban et al., 2018, Bisheban et al., 2019).

3. Adaptive and neural mechanisms

The adaptive-neuro layer enters in several distinct ways. In the quadrotor wind-rejection papers, each disturbance channel is modeled by a three-layer feedforward neural network,

$$\mathbb{R}^3 \times SO(3)\times (\mathbb{S}^2 \times SO(3))^n$$2

with estimated output

$$\mathbb{R}^3 \times SO(3)\times (\mathbb{S}^2 \times SO(3))^n$$3

One network estimates the disturbance force in the translational dynamics and another estimates the disturbance moment in the rotational dynamics. The position-network input is built from $$\mathbb{R}^3 \times SO(3)\times (\mathbb{S}^2 \times SO(3))^n$$4 and $$\mathbb{R}^3 \times SO(3)\times (\mathbb{S}^2 \times SO(3))^n$$5, while the attitude-network input uses $$\mathbb{R}^3 \times SO(3)\times (\mathbb{S}^2 \times SO(3))^n$$6 and $$\mathbb{R}^3 \times SO(3)\times (\mathbb{S}^2 \times SO(3))^n$$7 (Bisheban et al., 2019). The controller then inserts $$\mathbb{R}^3 \times SO(3)\times (\mathbb{S}^2 \times SO(3))^n$$8 into the intermediate vector

$$\mathbb{R}^3 \times SO(3)\times (\mathbb{S}^2 \times SO(3))^n$$9

that defines the desired thrust direction, and inserts $SO(3)$0 into the moment law $SO(3)$1 (Bisheban et al., 2018).

The neural weights are updated online by Lyapunov-derived laws. In the quadrotor formulations, the key filtered errors are

$SO(3)$2

and the update laws include learning-rate terms $SO(3)$3 and leakage terms $SO(3)$4, together with projection or clamping to enforce bounds on $SO(3)$5, $SO(3)$6, and $SO(3)$7 (Bisheban et al., 2018, Bisheban et al., 2019). The practical role of projection is explicit: it preserves bounded weights and yields bounds on the residual approximation term used in the Lyapunov derivative.

A second pattern is feature learning followed by online coefficient adaptation. In the Hamiltonian $SO(3)$8 controller, the disturbance is structured as

$SO(3)$9

where $\Delta_1,\Delta_2$0 is a learned matrix of nonlinear features and $\Delta_1,\Delta_2$1 is an unknown constant disturbance realization. The feature map $\Delta_1,\Delta_2$2 is identified offline by a Hamiltonian neural ODE trained from state-control trajectory data, while the online law adapts only the coefficients $\Delta_1,\Delta_2$3 through

$\Delta_1,\Delta_2$4

The disturbance compensation term is then

$\Delta_1,\Delta_2$5

inside an IDA-PBC controller (Duong et al., 2021). This architecture is explicitly not online retraining of the neural network; it is online adaptation of the last linear layer coefficients.

A third pattern appears in centralized multi-quadrotor transportation. There, adaptive-neuro geometric control is realized by coactively tuning model parameters and learning lumped disturbance dynamics in real time. The top-level payload controller augments geometric PD laws with RBF neural approximators $\Delta_1,\Delta_2$6, $\Delta_1,\Delta_2$7, adaptive estimates of the payload mass and inertia, and integral disturbance estimators for payload and cable disturbances (Gao et al., 2 Sep 2025).

These designs all inherit a template already visible in purely adaptive geometric control. On $\Delta_1,\Delta_2$8, for matched disturbances $\Delta_1,\Delta_2$9, the adaptive torque

$SE(3)$0

and the update law

$SE(3)$1

show how disturbance-cancellation terms are inserted additively into a geometric controller and tuned by Lyapunov cross-term cancellation (Kulumani et al., 2016). This precursor structure is the direct basis for later neural generalizations.

4. Stability guarantees and convergence claims

The stability claims in adaptive-neuro geometric control are shaped by the disturbance model and by the approximation mechanism. In the quadrotor wind-rejection work, the controller is analyzed on $SE(3)$2 with Lyapunov functions that combine translational errors, rotational errors, and neural-weight errors. The result is uniform ultimate boundedness of all tracking and weight errors, with an ultimate bound that can be abridged arbitrarily in one paper and reduced arbitrarily in the later experimental paper (Bisheban et al., 2018, Bisheban et al., 2019). The bounded residual arises from universal approximation error, from the bounded term $SE(3)$3 in the neural decomposition, and from the coupling between attitude and translation.

The $SE(3)$4 attitude controller with state inequality constraints yields a stronger statement under a more restrictive disturbance model. For the adaptive matched-disturbance case, the zero equilibrium of $SE(3)$5 is asymptotically stable, the parameter estimate remains bounded, and the state inequality constraints are satisfied, provided the design parameter $SE(3)$6 satisfies the explicit matrix-positivity condition given in the paper (Kulumani et al., 2016). In the nominal, disturbance-free case, the desired equilibrium is almost globally asymptotically stable and the other critical points are unstable.

The Hamiltonian $SE(3)$7 feature-learning formulation also gives convergence to zero under an exact-feature assumption. Theorem 1 states that, when the disturbance is exactly $SE(3)$8, the tracking errors $SE(3)$9 and $SE(3)$0 converge to zero and the parameter error $SE(3)$1 is Lyapunov-stable and uniformly bounded, within a region characterized by $SE(3)$2 and bounded angular-velocity error (Duong et al., 2021). The Lyapunov function is the shaped Hamiltonian plus a cross term and a parameter-error term.

The multi-quadrotor transportation formulation gives a different guarantee: Lyapunov-based adaptation guarantees bounded estimation errors without requiring either pre-training or the persistent excitation condition, and the full control system is proven to be stable in the sense of Lyapunov under certain preconditions (Gao et al., 2 Sep 2025). The resulting stability statement is practical rather than exact-convergence under arbitrary uncertainty.

A plausible implication is that the field exhibits a spectrum of guarantees. When the disturbance is matched or exactly represented in a learned feature space, asymptotic convergence of tracking errors is available. When neural universal approximation errors or unstructured aerodynamic effects remain, the standard conclusion is uniform ultimate boundedness rather than exact convergence.

5. Disturbance rejection, constraints, and implementation patterns

A major motivation for adaptive-neuro geometric control is disturbance rejection without an explicit disturbance model in the control law. The quadrotor wind papers present detailed aerodynamic models for validation, including quadratic drag $SE(3)$3, rotor thrust coefficients depending on inflow and advance ratios, blade flapping with thrust-direction change $SE(3)$4, and reaction torques $SE(3)$5. The controller, however, uses a simplified model with constant thrust coefficients and fixed thrust direction, so all wind-induced drag, thrust-direction variation, flapping, and torque variation are lumped into the unknown disturbances $SE(3)$6 to be learned online (Bisheban et al., 2018, Bisheban et al., 2019). This division between controller model and validation model is central to the disturbance-rejection claim.

Constraint handling is another recurrent theme. On $SE(3)$7, state inequality constraints are enforced by a logarithmic barrier

$SE(3)$8

so that the potential blows up as the forbidden cone boundary is approached. Because $SE(3)$9 is embedded directly in the Lyapunov function, safety and stability are established in the same argument (Kulumani et al., 2016). In a different register, the Hamiltonian $SO(3)$0 paper separates learning into an offline disturbance-model identification stage and an online adaptive control stage, which makes the implementation pipeline explicit: train the Hamiltonian neural ODE from multi-environment transition data, then adapt disturbance coefficients online using geometric errors (Duong et al., 2021).

The online workflow is most explicit in the quadrotor NN controllers. Per sampling instant, the implementation measures or estimates $SO(3)$1; computes $SO(3)$2; evaluates the disturbance estimates $SO(3)$3; constructs the computed attitude $SO(3)$4; computes total thrust $SO(3)$5 and control moment $SO(3)$6; allocates these to rotor thrusts; and updates the neural weights with projection if bounds are exceeded (Bisheban et al., 2018). In the transport formulation, the same real-time loop is layered over a centralized force-and-moment allocation from the payload level to the cable tensions and then to individual quadrotors (Gao et al., 2 Sep 2025).

The practical design space is therefore heterogeneous. Some controllers rely on fully online neural adaptation with no pretraining, some use offline feature learning plus online linear adaptation, and some combine parameter adaptation, disturbance integrators, and neural approximators in a single Lyapunov design.

6. Adjacent usages, misconceptions, and boundaries of the term

A recurrent misconception is that any neuro-adaptive controller with a structured error surface is geometric control in the same sense as $SO(3)$7 or $SO(3)$8 tracking. The recent multi-agent sliding-mode paper makes the distinction explicit: no manifold geometry is used explicitly, and the “geometric” aspects are limited to graph-theoretic structure and set-theoretic constraints in Euclidean space (Chaudhari et al., 29 Jul 2025). Its contribution is relevant to adaptive-neuro control and to restricted-potential-function design, but it is not a Lie-group geometric controller.

A second misconception is that “geometric” must always mean explicit $SO(3)$9 or $SE(3)$0 notation. The neuro-adaptive boundary force controller for dual one-link flexible arms does not use Lie-group attitude geometry, yet it is built around geometric constraints, boundary energy flow, and an energy/Lyapunov functional that combines kinetic and elastic terms with neural and adaptive estimators (Hejrati, 2022). The geometry there is that of a constrained mechanical system and of boundary control, not that of rigid-body rotation groups.

A third misconception is that the neural component is always an online universal approximator of the disturbance. The literature contains three different regimes: fixed offline-learned controllers combined with online non-parametric adaptive gains for task satisfaction under unknown dynamics; offline-learned disturbance features with online adaptation of feature coefficients; and fully online neural-weight adaptation inside geometric controllers (Verginis et al., 2021, Duong et al., 2021, Bisheban et al., 2019). The corresponding guarantees and assumptions are therefore not interchangeable.

This suggests that adaptive-neuro geometric control is best understood as a family of designs rather than a single formalism. In its narrowest and most established form, it denotes coordinate-free control on $SE(3)$1, $SE(3)$2, $SE(3)$3, or products with $SE(3)$4, augmented by adaptive and neural disturbance compensation and justified by Lyapunov analysis. In a broader adjacent literature, the same phrase can also refer to structured mechanical or task-specification-based controllers in which geometry enters through constraints, planning, or state-space structure rather than through a fully intrinsic manifold formulation.