The 1,067× Speedup: Finding the Perfect Servo Trajectory by Checking Just Six Options

Every modern factory floor hides a silent negotiation. A CNC machine must drill 4,000 holes in a sheet of aluminum in under a minute. A robotic arm must swing a payload to an exact point without overshooting. A lithography tool must position a silicon wafer to within nanometers. Each device does this through a servo system — the workhorse technology of industrial motion control, where a motor drives a tool along an axis to a target position with breathtaking speed and precision.

Behind every one of those movements sits a quiet mathematical compromise. Motion planners want speed; motor physics demands restraint. Too aggressive a command, and the actuator hits its torque, current, or voltage limits — the motor "saturates," tracking errors appear, and finished parts come out wrong. Too conservative, and the machine crawls, sacrificing productivity for safetytz. Engineers have long split the difference with a blunt instrument: pre-computed motion profiles that assume the motor is simpler than it actually is, then arbitrarily cap acceleration to keep it out of trouble.

A team from Mitsubishi Electric Research Laboratories and Mitsubishi Electric Corporation has now removed that compromise. In a paper presented at the 2026 IEEE Conference on Decision and Control, they describe a reference reshaping algorithm that guarantees a servo trajectory stays within its actuator's true physical limits — while running at frequencies approaching 100,000 times per second. The key to the breakthrough is not a faster computer, but a faster idea. The researchers proved that the optimal "reshaped" trajectory can always be found by checking a handful of candidate solutions, computed with closed-form algebra and one small eigenvalue problem. No iterative search. No root-finding. No guesswork. The exact optimal answer, every time, in deterministic time.

The practical payoff is staggering. The new algorithm runs 337× faster than the previous reference reshaping method and 1,067× faster than generic numerical optimization. That is the difference between a system capped at roughly 1 kHz of planning frequency and one that can approach 100 kHz — an order-of-magnitude-and-a-half leap that opens up servo control regimes that were previously unreachable.

The Science

The problem starts with a mismatch of abstractions. A servo system has two layers. A motion planner generates a reference trajectory — the path, velocity, and acceleration commands that should drive the tool to its target. Then a tracking controller regulates motor currents and voltages to make the motor follow that reference.

The tension is that the planner typically ignores motor physics. It outputs an "aggressive" trajectory designed purely for speedbones. The motor, meanwhile, obeys a hard set of physical constraints: maximum current $I_{\max }$, maximum voltage from the DC bus $V_{\max }$, and a torque capacity that shrinks as speed rises, because at high angular velocity the motor's back-EMF — the counter-electromotive force generated by its own rotation — eats into the voltage available to push current through the windings.

Feed an infeasible trajectory into the tracking controller and you get constraint violations, tracking drift, and scrap parts. The classic remedy — arbitrarily lowering the planner's allowed acceleration — wastes productivity for the sake of safety. What engineers really want is a reference trajectory that stays just inside the motor's true feasible region.

That is what reference reshaping does. The idea, introduced in earlier work by the same group (referenced in the paper as [8]), is to add an intermediate module — a "reference reshaper" — between the motion planner and the tracking controller. At each planning interval $\Delta$, the planner hands the reshaper a desired acceleration $a_{\mathrm{des},k+1}$. The reshaper then solves a small constrained optimization: find the motor speed ${\omega }_{k+1}$ whose implied acceleration $\frac{\eta }{\Delta }{\omega }_{k+1}-\frac{v_k}{\Delta }$ is as close as possible to the desired acceleration, subject to the constraint that the implied motor torque stays within the speed-dependent torque bound $\gamma {\tau }_{\mathrm{m}}({\omega }_{k+1})$.

{{image: lavish close-up of the architecture diagram here}}

The formulation is clean, but the execution was slow. The constraint boundary is defined by a nonlinear equation — the torque bound ${\tau }_{\mathrm{m}}(\omega ,\mathbf{\beta })$ is a piecewise function of speed, with different regimes depending on where along the motor's torque-speed curve the operation sits (see equation 3 in the paper). The previous algorithm solved that nonlinear equation with iterative root-finding, which is expensive, initialization-dependent, and only guaranteed to find one root. That kept the whole system capped at about 1 kHz.

The Mitsubishi team's contribution was to realize they didn't need to solve the constraint equation numerically at all. They could characterize its structure analytically.

What They Found

The core mathematical result is surprisingly elegant. The optimization problem is convex-feeling and the optimal solution must satisfy the Karush–Kuhn–Tucker (KKT) conditions — the standard necessary conditions for constrained optimality. From those conditions, the researchers showed that the optimal motor speed ${\omega }_{k+1}^{\ast }$ must be one of only two things: either a stationary point of the objective function (which has a closed-form expression), or a point on the boundary of the feasible region.

The boundary points are the crux. The feasible region is bounded by the torque constraint, which lives on the curve $\gamma {\tau }_{\mathrm{m}}(\omega )=\frac{J}{\Delta }\omega -\frac{J{v}_k}{\Delta \eta }$ (and its mirror image). Because the torque bound ${\tau }_{\mathrm{m}}(\omega ,\mathbf{\beta })$ is piecewise — constant at low speed, then monotonically decreasing through a complex middle regime, then decreasing as $1/\omega$ at high speed — the boundary equation splits into at most three branchesasia (equations 7–9 in the paper).

The first branch (constant torque region) and the third branch (high-speed region) yield boundary points in closed form — simple linear or square-root expressions. The middle branch is the hard one. There, the constraint becomes a sixth-degree polynomial in $\omega$, and by the Abel–Ruffini theorem, a general sixth-degree polynomial has no closed-form roots.

Here is where the mathematics earns its keep. Instead of abandoning the analytical path, the researchers converted the polynomial's roots into the eigenvalues of a $6\times 6$ companion matrix $\mathbf{C}(\mathbf{\beta })$. Real eigenvalues of this small matrix give exactly the constraint boundary points in the middle branch, and computing eigenvalues of a $6\times 6$ matrix is a cheap, deterministic, well-conditioned operation — no iteration, no initial guess, no risk of missing a root. The complete algorithm enumerates the finite candidate set of at most six boundary points plus the stationary point, checks which ones are feasible, and picks the best. Theorem 1 in the paper proves this recovers the exact optimal solution of the original problem.

The speedup is dramatic. The paper reports a 337× improvement over the prior reference reshaping approach and 1,067× over numerical optimization, pushing the planning frequency from roughly 1 kHz toward 100 kHz.

There is a second, more practical contribution. When the reshaper modifies an aggressive command, it can introduce a subtle kinematic mismatch — the reshaped acceleration profile differs from the desired one, so the tool may accumulate a small position error, falling short of or overshooting the target $p_{\mathrm{T}}$. The authors introduced a real-time trajectory compensator that exploits the symmetry of time-optimal trajectories: when the desired and feasible trajectories are "time-symmetric," the compensator can insert a constant-velocity phase — a brief cruise at fixed speed — to recover the position deviation online, without ever pushing the torque back over its limit.

Computational speedup of the proposed reference reshaping algorithm

Label	Value
Numerical optimization	1,067
Prior reference reshaping	337
Proposed method	1

Source: Lu et al. (2026). Proposed method normalized to 1×; prior RR = 337×; numerical optimization = 1067×.

Here is how the three approaches compare in compute cost, compiled from the paper's reported performance numbers. The proposed method sits at the bottom of the scale, orders of magnitude below the alternatives.

Why This Changes Things

To appreciate what 100 kHz planning actually means, consider the landscape of industrial motion control. A 10 kHz control loop is already considered fast. Many industrial servos run their planning stage at a few hundred Hz to a few kHz precisely because the mathematically principled optimizers are too slow. The new algorithm removes that constraint. At 100 kHz, the reshaper can respond to changing conditions — a load change, an interference from a neighboring axis, a re-plan mid-motion — with latency measured in microseconds.

The deeper significance is philosophical. The servo control community has long had two competing ways of guaranteeing actuator feasibility. Trajectory optimization — solving the full optimal control problem with actuator dynamics baked in — is principled but computationally brutalanced to deploy in real time. Reference/command governors — which enforce constraints based on closed-loop system dynamics — are fast but rely on conservative, pre-computed safety sets that don't capture the full physics. The reference reshaper splits the difference: it operates directly on the desired trajectory (like a planner) while enforcing the motor's true torque-speed envelope derived from physics (like a governor). The paper frames this explicitly: the reshaper "combines elements of trajectory optimization and reference governing, while remaining specifically designed for high-frequency real-time implementation."

There is also a hardware angle. The torque bound ${\tau }_{\mathrm{m}}(\omega ,\mathbf{\beta })$ is derived from max torque per ampere (MTPA) control, the strategy that squeezes maximum torque out of a given current by optimally splitting it between the motor's d- and q-axes. The feasible envelope depends on real motor parameters — inductances $L_{\mathrm{d}},L_{\mathrm{q}}$, permanent-magnet flux ${\Phi }_{\mathrm{pm}}$, pole pairs $p$, inertia $J$, and the voltage and current limits $V_{\max },I_{\max }$. By reshaping against this physics-derived envelope rather than a conservative S-curve, the machine is allowed to run closer to its true performance edge. Productivity gains come not from a faster motor, but from not wasting the motor you already have.

{{image: torque-bound figure here}}

The torque-speed envelope in Figure 2 shows the geometry of the problem. At low speed, the motor can deliver its full constant torque. Past the corner speed ${\omega }_{\mathrm{r}}$, available torque decays as back-EMF eats the voltage budget. The reshaper keeps the operating point inside that red feasible region at every instant.

What's Next

The paper is honest about its scope. The system model is one-dimensional — a single work tool moving along one axis, driven by a surface-mounted permanent magnet synchronous motor (SPMSM). Real machines are multi-axis, with coupling between axeshare, gravity loads, friction, and compliance in the mechanical drivetrain that the model does not capture. Extending the analytical candidate-set approach to coupled multi-axis systems is an obvious and open challenge.

There are also practical deployment questions that only industrial testing will answer. The torque-speed bound depends on motor parameters that drift with temperature — winding resistance $R$ rises as the motor heats, and permanent-magnet flux ${\Phi }_{\mathrm{pm}}$ weakens slightly with temperature. The paper does not address online parameter identification or the robustness of the candidate-set approach to parameter error. Whether the 337× advantage holds up in a hot, vibrating, packet-lossy factory environment remains to be seen.

But the trajectory of the work is unmistakable. The authors proved that an optimal control problem, long thought to require iterative numerical solution, has a finite, closed-form describable solution set. That is the kind of result that resets what engineers assume is possible in real time. The servo field has been nibbling at the 1 kHz planning barrier for years; this clears it by two orders of magnitude.

The broader lesson extends beyond servos. Many real-time optimization problems — in robotics, power electronics, autonomous vehicles, chemical process control — are bottlenecked not by the size of the problem but by the assumption that it must be solved iteratively. When a problem's structure admits a finite candidate set, determinism, and closed-form enumeration, "real-time" stops being a constraint and becomes a feature. The Mitsubishi team has shown, with rigor and a working proof, that sometimes the fastest algorithm is not the one with the better solver — it is the one that realizes there is nothing to iterate over at all.