The Math Trick That Makes Robots Easier — and Provably Best — to Control

Ask a control engineer how "easy to control" a robot is, and you'll likely get a shrug. Ease of control is one of those qualities everybody wants and almost nobody can define — it seems to depend on the application, the operator, even the mood of the moment. But for a guidance robot steering a visually impaired user along a sidewalk, "easy to control" has a very concrete meaning: the user should be able to push the handle forward and have the robot glide along the path, but the robot should resist being shoved sideways into traffic. The directions in which the system moves easily, and the directions in which it resists, are a design choice.

A team of Japanese researchers has just made that choice dramatically easier to compute. In a new paper, Koju Nishimoto, Yuki Onishi, Riku Funada, and Mitsuji Sampei show that shaping a system's "controllability Gramian" — the mathematical object that decides how cheaply you can move the state of a system in each direction — can be turned into a convex optimization problem, the kind computers solve quickly and reliably (Nishimoto et al., 2026). The previous state of the art could get stuck in local minima, like a hiker settling for a small hill instead of climbing to the true peak. The new method guarantees the globally best answer every time, and it does so using a distance borrowed from a surprising corner of mathematics: optimal transport, the same machinery that powers modern image generation and self-driving-car trajectory planning.

The practical payoff is real. In numerical experiments, the method lets an engineer specify, say, "make the robot easy to pull forward but hard to push sideways," and the computer hands back a controller that does exactly that — with a hard, provable guarantee that no better controller exists.

The Science

To understand what's new, you first need to meet the controllability Gramian. For any linear system with a stable "engine" — think of it as the dynamics matrix $\mathbf{A}$ — the Gramian $\mathbf{P}$ is a symmetric matrix that encodes how much input energy it takes to reach any given state. If you want to move the state to some point ${\mathbf{x}}_0$, the minimum energy required is ${\mathbf{x}}_0^{\top }{\mathbf{P}}^{-1}{\mathbf{x}}_0$. The set of states reachable with unit input energy, $\{\mathbf{x}\mid {\mathbf{x}}^{\top }{\mathbf{P}}^{-1}\mathbf{x}\le 1\}$, is an ellipsoid, and the shape of that ellipsoid is the whole story: directions where the ellipsoid is long are directions the system moves easily; directions where it is thin are directions the system resists.

The Gramian is not fixed by fate. In systems driven by an external operator — a human gripping a robot's handle — there are two inputs: the internal control $\mathbf{u}$, which the engineer designs, and the exogenous input $\mathbf{v}$, which is the human's will. By choosing a feedback gain $\mathbf{K}$ (so the internal controller reacts to the state), the engineer reshapes the Gramian that the exogenous input "sees." This is the essence of "Gramian shaping": pick a desired Gramian ${\mathbf{P}}_d$ that encodes the operability you want, and tune $\mathbf{K}$ to pull the actual Gramian $\mathbf{P}$ as close to ${\mathbf{P}}_d$ as possible.

The catch is how you measure "close." The previous approach used the affine-invariant Riemannian (AIR) distance, which respects the curved geometry of positive definite matrices. But AIR-based shaping produced a nonconvex optimization problem. Convex problems have a wonderful property: any solution you find is guaranteed to be the global one. Nonconvex problems can trap you in local minima, make results depend on your starting guess, and generally behave like a foggy maze. Worse, the AIR distance resists physical interpretation — it's a number, but it's hard to say what that number means.

The authors' move is to swap in the Bures–Wasserstein (BW) distance. Defined on positive semidefinite matrices $\mathbf{X},\mathbf{Y}$, it takes the form

$$d_{\mathrm{BW}}(\mathbf{X},\mathbf{Y})={\left(\mathrm{tr}\left(\mathbf{X}+\mathbf{Y}-2{\left({\mathbf{Y}}^{\frac{1}{2}}\mathbf{X}{\mathbf{Y}}^{\frac{1}{2}}\right)}^{\frac{1}{2}}\right)\right)}^{\frac{1}{2}}.$$

That formula looks forbidding, but its meaning is lovely: the BW distance coincides exactly with the 2-Wasserstein distance between two Gaussian distributions with the same mean. In other words, measuring the gap between two Gramians is the same as asking how much "work" it takes to rearrange one cloud of points into another cloud of points — the very question behind optimal transport's modern renaissance in machine learning.

The paper's first theorem is the hinge on which everything turns: the squared BW distance to a fixed desired Gramian is strictly convex on the set of symmetric positive definite matrices. The proof is elegant. The critical ingredient is that the trace of a matrix square root, $\mathrm{tr}({\mathbf{X}}^{1/2})$, is strictly concave — a known fact the authors deploy like a chess piece. Strict concavity of the square-root trace flips the inequality exactly the right way, and out pops strict convexity of the objective (Nishimoto et al., 2026). From there, uniqueness follows immediately: if a solution exists, there is only one, and whatever a convex solver returns is it.

What They Found

The first result is the convexity guarantee itself. Where the AIR approach was nonconvex, BW-based shaping produces a problem with a strictly convex objective and an affine feasible set, so any converged solution of a standard solver is provably the global optimum. This is not a minor nicety; it transforms Gramian shaping from an art (dependent on initialization, prone to local traps) into a science.

The second contribution is computational. The authors show the shaping problem can be recast as a semidefinite program (SDP) using an LMI representation of the BW distance. Through a change of variables $\mathbf{E}=\mathbf{K}\mathbf{P}$, the Lyapunov constraint that ties together the Gramian and the feedback gain becomes a linear equation in $\mathbf{P}$ and $\mathbf{E}$, and the BW distance becomes an inequality on a block matrix

$$\left[\begin{array}{cc}\mathbf{P}& \mathbf{U}\\ {\mathbf{U}}^{\top }& {\mathbf{P}}_d\end{array}\right]⪰0.$$

Once the problem lives in the SDP world, it inherits two huge advantages. First, mature, fast solvers exist, so the computation is tractable even as the system grows. Second — and this is where the design flexibility leaps forward — any constraint expressible as an LMI can be bolted on. Want an upper bound on certain entries of the Gramian? That's an LMI. Want a norm bound on the internal control effort? Also an LMI. The framework stops being a fixed recipe and becomes a toolbox.

The numerical experiments drive the point home. The authors consider a guidance robot — the kind that steers a visually impaired user along a track, shown in

— navigating a path with an exogenous input from the user's handle. In one set of cases, the exogenous input gets reflected in the state but causes the robot to drift off its reference path. The key result: by adding a directional LMI constraint, the designer can suppress that deviation entirely, producing a trajectory that hugs the reference path (

). This is anisotropy in action — making the robot easy to control in one direction and stiff in another, exactly as the design brief demanded.

The third finding connects the framework to classical control theory in a way that feels almost like a discovery about the mathematics rather than an engineering trick. When the exogenous input is modeled as zero-mean Gaussian white noise — the standard model for unpredictable disturbances — BW-based Gramian shaping turns out to be closely tied to $H_2$ control, the workhorse of optimal control that minimizes the expected energy of the output (Nishimoto et al., 2026). The link is precise: as the desired Gramian ${\mathbf{P}}_d$ shrinks toward zero, the optimal stationary covariance produced by the method converges to the covariance produced by $H_2$ control.

plots the difference $\Vert {\mathbf{P}}^{\ast }-{\mathbf{P}}_{H_2}{\Vert }_F$ against the scaling factor $\alpha$, and the curve falls monotonically toward zero.

This is a genuinely striking unification. $H_2$ control is one of the most successful and widely deployed tools in all of engineering — it underpins everything from airplane autopilots to suspension design. Optimal transport is the darling of modern machine learning. The paper shows they are not distant cousins but two ends of the same spectrum, with Gramian shaping as the bridge. In the language of the paper, $H_2$ control "can be interpreted as a special case of BW-based Gramian shaping," which itself is a special case of optimal transport.

shows the state sample distributions under the proposed method for different $\alpha$, narrowing toward the $H_2$ distribution as $\alpha$ decreases.

Why This Changes Things

Let's be clear about what this paper does not claim. It does not invent a new physical actuator or a new sensor. It works on a far more subtle level: it changes what a designer can prove about the controller they build. Before this work, shaping the controllability Gramian to hit a desired operability profile meant solving a problem whose answer might depend on where you started. You could tune, iterate, and hope. You could not certify. The BW reformulation hands the designer a guarantee: the controller you get is the best one possible, full stop, and if you add constraints, they are honored exactly.

The physical meaning is worth dwelling on. The BW distance's identity with the Wasserstein distance between Gaussians gives Gramian shaping a concrete interpretation that the AIR distance never had. If you imagine the state distribution of the system (under Gaussian noise) as a cloud of particles, then the "distance" between your current Gramian and your desired one is literally the cost of moving those particles into their ideal arrangement. Designing a controller becomes a problem in logistics: how much work does it take to reshape this cloud into that cloud? That is a sentence that could be said by someone studying economics or fluid dynamics, and its arrival in control theory is a small intellectual pleasure.

The practical stakes are human. Consider the guidance robot in

. Its user is visually impaired, gripping a handle, being pulled along a track. The design problem is not "make the robot fast" but "make the robot follow the user's intent without letting a misstep become a catastrophe." The engineer wants the robot easy to pull forward along the path — cheap in input energy — but resistant to sideways shoves that could pull the user off the curb. That is anisotropic controllability, and it is precisely what the method delivers. The same logic applies to manipulators with direct teaching (where a worker physically drags a robot arm to program it) and to any human-interactive system where operability and safety must be balanced. Impedance and admittance control, the classical tools for shaping human-robot interaction, mainly prescribe local input-output relations; the Gramian approach designs directional reachability over the entire state space at once.

There is also an efficiency story. Nonconvex optimization in the AIR framework was not just fragile; it was slow, and it got slower and less reliable as systems scaled. SDP formulations, by contrast, are served up fast by off-the-shelf solvers. For an engineer iterating on a design, the difference between "wait, I need to re-tune my initial guess" and "the convex solver returned the global optimum in a fraction of a second" is the difference between friction and flow.

What's Next

The natural next steps are both theoretical and applied. On the theory side, the connection to $H_2$ control invites a fuller exploration of where on the BW-to-$H_2$ spectrum the interesting designs live. The paper shows that as ${\mathbf{P}}_d\to 0$, BW shaping converges to $H_2$ control — but the interesting regime is the opposite one, where the desired Gramian encodes deliberate anisotropy, spatial preferences, and human ergonomics that classical $H_2$ theory cannot express. Mapping this territory, and understanding what kinds of operability profiles are achievable (and at what cost in control effort), is an open and fertile question.

The authors are honest about the caveats. The framework, as derived, lives firmly in linear systems. The transition to nonlinear systems — where the state distribution under noise is no longer Gaussian, and where the Wasserstein machinery gets substantially harder — is a genuine gap. Guidance robots and direct-teaching manipulators are themselves increasingly nonlinear in their full-fidelity models. The convexity guarantee, so clean in the linear setting, will need new ideas to survive contact with nonlinear reality. There is also the matter of robustness: the method designs for a nominal model, and how gracefully the shaping degrades under model uncertainty is not yet addressed.

But the direction is unmistakably forward. By proving that controllability Gramian shaping is a convex problem solvable as an SDP, and by exposing its lineage to both $H_2$ control and optimal transport, this paper does two things at once: it makes a practical tool more trustworthy, and it shows that two great mathematical traditions in control and in machine learning were, all along, branches of the same tree. The next time a robot resists being shoved off a curb, or an arm yields exactly as a worker drags it through a taught motion, part of the credit may trace back to a distance function borrowed from the flow of particles.