When Symmetry Breaks: How a New Mathematical Framework Makes Drones Smarter

An aerial vehicle flying through the air—whether a drone, helicopter, or bird—should, in theory, behave the same no matter where it is or which way it’s facing. That’s symmetry: the idea that certain transformations—like shifting position or rotating in place—don’t change how a system evolves. For decades, engineers have exploited such symmetries to simplify control problems, design better navigation systems, and build more stable robots. But real-world physics rarely cooperates. Gravity pulls down. Wind pushes sideways. These forces break perfect symmetry, making classical symmetry-based methods fail or become overly complex.

Now, a new mathematical framework called weak invariance redefines what it means for a system to be symmetric—not by insisting on perfection, but by allowing asymmetry to be structured. In a 2026 paper by Jake Welde, Riley Link, and Pieter van Goor, the authors show that even when symmetry is broken, the “residual” asymmetry—the part that doesn’t fit—can itself be modeled as a control system on the symmetry group. This insight unlocks a broader class of systems that can still benefit from symmetry-inspired simplifications, even when traditional symmetry fails.

The most striking result? An aerial vehicle under gravity—a classic example of broken symmetry—admits a nine-dimensional weak symmetry, nearly doubling the size of its previously known four-dimensional classical symmetry. This isn’t just a mathematical curiosity. It means engineers can now factor out more complexity from the system’s error dynamics, reducing the effective dimensionality of the problem and enabling more efficient, robust control algorithms.

The Science

At the heart of this work is a shift in perspective: instead of asking whether a system is symmetric, ask how it fails to be symmetric. The authors formalize this using the language of differential geometry and Lie groups—mathematical structures that describe continuous symmetries like rotation and translation.

Consider a control system with state $x$ evolving on a manifold $M$, influenced by inputs $u$. A Lie group $G$ acts on $M$ via a map $\Phi: G \times M \to M$, representing transformations like rotation or translation. Classical symmetry (or strong invariance) requires that the dynamics commute with this action:

$$\mathrm{d}{\Phi }_{g^{-1}}\circ f_u\circ {\Phi }_g=f_u$$

This means that transforming the state and then applying the dynamics is the same as applying the dynamics and then transforming—essentially, the system behaves the same before and after the transformation.

But when external forces like gravity are present, this equality fails. The authors introduce a residual $\Delta(g,u)$ that quantifies this failure:

$$\Delta (g,u)=\mathrm{d}{\Phi }_{g^{-1}}\circ f_u\circ {\Phi }_g-f_u$$

If the residual is zero, the system is strongly invariant. If the residual is non-zero but lies in the space of vertical vector fields (tangent to group orbits), the system is partially invariant—a known generalization. But Welde, Link, and van Goor go further. They define weak invariance as the case where the residual lies in the span of the infinitesimal generators of the group action, $\mathfrak{g}(M)$. This is a stricter condition than partial invariance but far more flexible than strong invariance.

Crucially, they prove that weak invariance is equivalent to the residual being generated by another control system evolving on the symmetry group $G$. This means the asymmetry isn’t arbitrary—it follows its own structured dynamics.

The authors also show that weakly invariant systems admit a cascade decomposition: the full system can be split into two parts, one evolving on the symmetry group $G$ and the other on the quotient space $M/G$. The subsystem on $G$ is group affine—a class of systems with favorable properties like state-independent error dynamics, which are crucial for stable observer design.

What They Found

The paper’s central technical contributions are threefold:

1. Weak invariance generalizes both classical symmetry and group affine systems. While classical symmetry requires $\Delta = 0$, and partial symmetry allows $\Delta$ to be any vertical vector field, weak invariance restricts $\Delta$ to be in $\mathfrak{g}(M)$—the set of infinitesimal generators. This seemingly small change has profound implications. The authors prove that any weakly invariant system admits a cascade decomposition where the $G$-subsystem is group affine. This bridges two previously distinct classes of systems: invariant systems and group affine systems (which include left-invariant systems as a special case).

2. Weak symmetry enables greater dimensionality reduction in error dynamics. When the residual is autonomous (i.e., independent of control inputs), the authors show that the symmetry group can be factored out of the error dynamics between two trajectories. This leads to a reduced-order error system whose dimension is lower by $\dim G$. For classical symmetries, this reduction is already powerful; for weak symmetries, it applies to a much broader class of systems.

3. A real-world system—the aerial vehicle under gravity—exhibits a larger weak symmetry than previously known. The authors analyze a standard model of a rigid body in 3D space under gravity. Classically, this system has a 4D symmetry: translations in $x$ and $y$, and rotations around the vertical ($z$) axis. But under weak invariance, they identify a 9D symmetry group, strictly containing the classical one. This larger symmetry arises because the residual due to gravity can itself be modeled as evolving on a Lie group.

Symmetry Group Dimension for Aerial Vehicle

Label	Value
Classical (Strong) Symmetry	4
Weak Symmetry	9

The weak symmetry group is 125% larger than the classical symmetry group, enabling greater dimensionality reduction in control and estimation.

This expansion—from 4D to 9D—is not just a mathematical artifact. It means that more of the system’s structure can be exploited for control and estimation. The larger the symmetry group, the more dimensionality can be factored out, leading to simpler, more robust algorithms.

The authors also derive conditions under which a subgroup of a weak symmetry is itself a weak symmetry, and show how to extract the largest strong symmetry from a given weak symmetry. This provides a way to “upgrade” weak symmetries when possible, or to understand the limits of classical symmetry in a given system.

Why This Changes Things

Symmetry is not just a theoretical elegance—it’s a practical tool. In robotics, aerospace, and autonomous systems, symmetry-based methods have enabled:

• Invariant observers that maintain consistent estimation performance regardless of orientation or position [13, 14].
• Equivariant neural networks that generalize better in reinforcement learning by encoding physical symmetries [18].
• Reduced-order models that speed up simulation and control design by factoring out redundant dimensions [19, 20].

But these methods have been limited to systems with exact symmetry. Real-world systems—drones in wind, cars on hills, robots manipulating objects—are rarely perfectly symmetric. Engineers have had to either ignore symmetry altogether or approximate the system as symmetric, risking performance loss or instability.

Weak invariance changes this. It provides a rigorous framework for systems that are “almost symmetric” in a structured way. The key insight is that asymmetry can be dynamic, not just static. Gravity doesn’t just break symmetry—it does so in a way that can be modeled as its own control system on the symmetry group. This structured failure is what makes weak invariance so powerful.

Consider the implications for drone navigation. A quadcopter flying indoors must estimate its position and orientation while compensating for gravity and aerodynamic disturbances. Traditional filters like the Kalman filter struggle with the nonlinearities and coupling in such systems. In contrast, invariant filters exploit symmetry to design error dynamics that are independent of the current state, leading to more predictable and stable performance [22].

But gravity breaks the full SE(3) symmetry of free flight, limiting the applicability of these methods. With weak invariance, engineers can now design filters that account for gravity’s effect within the symmetry framework itself. The 9D weak symmetry identified in the paper suggests that more of the system’s structure can be preserved in the error dynamics, potentially leading to filters with faster convergence and better robustness.

Moreover, weak invariance opens the door to learning structured asymmetries from data. In data-driven control, one could learn not just the dynamics, but also the residual symmetry-breaking terms as a separate system on the group. This could lead to more interpretable, generalizable models—imagine a robot learning that “down” is special not as an arbitrary bias, but as a structured deviation from symmetry.

Hierarchy of Symmetry Classes

Label	Value
Strong Invariance	0
Partial Invariance	1
Weak Invariance	2
General Nonlinear	3

0 Strong Invariance

Weak invariance occupies a middle ground, generalizing strong invariance while preserving useful structural properties.

The framework also unifies previously disconnected ideas. Group affine systems, which have been studied for their favorable filtering properties [30, 31], are now seen as a special case of weakly invariant systems. This suggests that the success of group affine models in practice may stem from their ability to capture structured asymmetry—even when the full system isn’t classically symmetric.

What’s Next

The paper lays a theoretical foundation, but the real impact will come from applications. Several open questions and next steps emerge:

• Can weak invariance be detected automatically in real systems? For classical symmetry, one can check invariance by testing whether the dynamics commute with group actions. For weak invariance, the challenge is to identify not just the group, but the residual system on the group. This may require new algorithms for symmetry discovery in data-driven settings.

• How does weak invariance interact with learning? In reinforcement learning, symmetry is often encoded as an inductive bias to improve sample efficiency [18]. Could weak invariance be used to design policies that learn to compensate for structured disturbances like gravity or friction, rather than treating them as noise?

• Can weak symmetry be exploited for motion planning? If the error dynamics can be reduced by $\dim G$, planning in the reduced space could be faster and more reliable. This could be especially useful in high-dimensional systems like humanoid robots or autonomous vehicles.

• What are the limits of weak invariance? The paper assumes free and proper group actions, which may not hold in systems with singularities (e.g., gimbal lock in attitude control). Extending the framework to local or singular actions would broaden its applicability.

One particularly promising direction is observer design for weakly invariant systems. The authors show that when the residual is autonomous, the group error dynamics are state-independent. This is a hallmark of invariant filters, which have been shown to outperform standard EKFs in navigation tasks [25, 28]. A natural next step is to derive a full observer for the aerial vehicle example, leveraging the 9D weak symmetry to achieve better estimation performance under gravity.

Another frontier is control design. The cascade decomposition suggests that one could design controllers that stabilize the group-affine subsystem first, then the base system. This hierarchical approach could simplify control of complex mechanical systems.

Finally, the framework may have implications beyond engineering. In neuroscience, for example, animal locomotion often exhibits approximate symmetries—think of a fish swimming forward, where left-right symmetry is broken by turns. Weak invariance could provide a language to describe how such systems maintain robustness despite asymmetry.

The story of symmetry in control is evolving. From rigid invariance to structured flexibility, weak invariance represents a shift from perfection to practicality. It acknowledges that the world is messy—but that even in the mess, there is structure waiting to be exploited.