Research analysis · Control and closed loops

Spiking control that stabilizes a robot's orbit, with a stability proof

A team at Sirius University gives event-based spiking feedback control its first stability guarantees for periodic orbits of underactuated robots. By extending a greedy Lyapunov-based design from linear time-invariant to periodic linearized dynamics, they construct a controller that fires constant-amplitude spikes only when each spike provably decreases a Lyapunov function, and they validate it on a physical butterfly robot tracking a 2.38-second orbit.

Source: Spiking Neuromorphic Control for Periodic Orbits Stabilization in Underactuated Robots: A Greedy Lyapunov-Based Approach, arXiv:2610.37310, 2026. Primary source. Read the full HTML of the preprint, including the control design and experimental sections.

What the work claims

This is a control-theory paper with a physical validation, and its claim is a guarantee rather than a performance number: feedback built from sums of constant-amplitude spikes, the currency of spiking neural networks, can stabilize periodic orbits of underactuated mechanical systems, and the closed loop can be certified. The technical step is extending a greedy Lyapunov-based spiking design, previously proved only for linear time-invariant (LTI) systems, to the linear time-varying systems with periodic coefficients that arise when robot dynamics are linearized transversely to a periodic orbit.1

The demonstrated result is practical orbital stability of a reference trajectory on the butterfly robot, a benchmark with two degrees of freedom, one control input (hence underactuated), a figure-eight link driven by a DC motor, and a ball rolling passively along the link's surface. The reference orbit has a period of 2.38 s. The controller was tested in a series of physical experiments with spike amplitudes of 5, 6, 10, 12, and 20 x 10^-3, with the presented results at an amplitude of 0.02; for all tested amplitudes the transverse coordinates converged to a neighborhood of the origin, which is exactly what practical stability means: convergence to a tube around the target, not asymptotically onto it.1

How it works

The construction has three layers. First, transverse linearization: near a periodic orbit, the robot's dynamics are rewritten in coordinates that measure only deviation transverse to the orbit, yielding a linear system with periodic coefficients whose trivial solution corresponds to perfect orbit tracking. Stabilizing the orbit becomes stabilizing that periodic linear time-varying system. Second, the spike mechanism: the control input is restricted to a sum of fixed-amplitude spikes arriving at event times, so between spikes the system evolves continuously and at each spike its state jumps discontinuously. The closed loop is therefore a hybrid system with periodic linear flow and impulsive corrections.

Third, the greedy firing rule, which is the paper's engine. The authors build a quadratic Lyapunov function V(tau, xi) = xi^T P(tau) xi, where P(tau) is the stabilizing periodic solution of a matrix Riccati differential equation (solved numerically, with weighting matrix Q = diag{20, 1, 1} and coefficient R = 2 x 10^4, then interpolated with B-splines for real-time use). A spike fires only when it provably decreases V, chosen greedily from the available spike directions; between spikes V decays under the periodic dynamics. Monotonicity of V across both flow and jumps establishes practical stability of the orbit. The controller is event-triggered in a strict sense: nothing happens between informative moments, and each action is a certificate-carrying discrete kick rather than a continuous torque.1

Where a skeptic should push

The most load-bearing assumption is that the linearized periodic model is valid where the robot actually operates. The guarantee lives in a neighborhood of the reference orbit; the paper reports convergence to a neighborhood (practical stability), not exact tracking, and the size of that neighborhood on the hardware is set by unmodeled contact forces, non-ideal motor dynamics, vision noise, and measurement delays, all of which the authors name as the cause of frequent spike triggering. That last observation deserves weight: the event-based ideal is sparse, information-driven firing, yet on the physical robot spikes arrive densely enough that the authors low-pass filter the control signal at 20 Hz to display it. When disturbances dominate, the purported parsimony of spiking control can collapse toward a high-rate pulse train, and with it much of the energy argument for neuromorphic control.

Second, no neuromorphic hardware is in this loop. The spiking controller is implemented as a computed event-based control law: the Riccati solution and firing conditions are interpolated and evaluated in a conventional real-time control loop, with an encoder and a computer-vision system supplying state estimates. That is a legitimate design choice, but the paper demonstrates that a spiking control policy can be certified and executed, not that a neuromorphic chip saves energy doing it. Third, the demonstration is a single 2-degree-of-freedom testbed tracking one orbit; the Riccati machinery scales awkwardly with state dimension, and robustness to model error enters only through the practical-stability margin, not through a designed-in robustness guarantee. The theory is a real extension; the hardware story is a feasibility demonstration.

Closed-loop spike control meets living tissue

The non-obvious implication for organoid intelligence is that this paper supplies the missing theoretical vocabulary for closed-loop stimulation of living neural tissue. Every serious proposal for training organoids, optogenetic or electrical, eventually confronts the same question: when, and how hard, should the machine touch the culture? Continuous feedback is biologically implausible and energetically wasteful; naive periodic stimulation ignores the state of the tissue. The greedy Lyapunov rule is a principled third option: define a certificate function over the culture's decoded state, and stimulate only when a spike provably decreases it. Reframed, a stimulation pulse is an impulsive control input to a hybrid system, and the question "does closed-loop training converge?" becomes "does a certificate decrease across spike events?" That reframing is exactly what the field needs before it can distinguish a training protocol that works from one that merely perturbs.

The certificate machinery also clarifies what would have to be true biologically. The proof needs a stabilizing periodic solution of a Riccati equation over linearized transverse dynamics, which presumes a model of the tissue's response to stimulation that is linearizable, time-invariant or slowly varying, and identifiable. None of these holds cleanly for cortical organoids, whose effective connectivity drifts over days and whose response to optical stimulation is state-dependent and history-dependent. The honest transfer is therefore the design pattern, not the theorem: define a Lyapunov-like objective over decoded dish state, trigger stimulation greedily, and monitor the objective's trajectory as the empirical certificate. Several published closed-loop organoid training schemes already do something like this with heuristic reward signals; this paper shows the certified version exists in the control literature and spells out which assumptions buy the guarantee.

The threat is the paper's own hardware result. On a system far better characterized than any culture, a robot with a known Lagrangian and clean sensors, disturbances and delays still drove the controller into dense triggering and a practical, not exact, stability margin. A cortical organoid presents orders of magnitude more unmodeled dynamics, sensor noise, and delay. Two consequences follow. For the optimists: expect the same collapse of sparsity; an event-driven stimulation protocol that fires only on informativeness will, on living tissue, fire often, and its energy and phototoxicity budgets must be sized for that reality, especially for optogenetic actuation where light dose is a biological stressor. For the skeptics: the certificate view gives a concrete failure detector. If the monitored Lyapunov-like quantity stops decreasing across stimulation events, the protocol is not training the tissue; it is harassing it. Building that alarm into closed-loop biological computing rigs, and publishing when it trips, would raise the field's credibility more than another demonstration of apparent learning.

The bottom line

Established: a greedy Lyapunov-based spiking controller, extended from LTI to periodic linearized transverse dynamics, achieves practical orbital stability of a butterfly robot's 2.38-second periodic orbit on physical hardware across spike amplitudes from 5 to 20 x 10^-3, with a proof that the hybrid closed loop's Lyapunov function decreases across flow and jumps. Not established: energy savings on neuromorphic hardware (none was used), sparsity of firing under realistic disturbances (the physical experiment triggered frequently), or robustness guarantees beyond a practical-stability neighborhood around a single validated orbit. What would confirm the broader program: deployment on an actual neuromorphic chip with an energy comparison against a continuous controller, and extension to systems where the linearized model is deliberately imperfect. What would weaken it: examples where greedy firing provably destabilizes outside the validity neighborhood. For organoid intelligence, the durable contribution is conceptual: closed-loop stimulation of living tissue is hybrid impulsive control, certificates over spike-triggered decreases are the right way to think about convergence, and the robot's dense-triggering behavior under noise is the cautionary precedent that closed-loop optogenetic training budgets should assume.

Frequently asked questions

What is transverse linearization?

A coordinate change near a periodic orbit that keeps only deviations perpendicular to the orbit itself. Tracking the orbit becomes stabilizing the zero solution of a linear system with periodic coefficients, which standard stability theory can handle even though the original robot dynamics are nonlinear.

What does the greedy Lyapunov rule do?

It constructs a quadratic energy-like function V(tau, xi) = xi^T P(tau) xi from the periodic solution of a Riccati equation, then fires a constant-amplitude spike only when that spike decreases V. Between spikes the periodic dynamics already decay V, so V decreases everywhere, which proves practical stability of the closed loop.

What is practical stability, as opposed to asymptotic stability?

Practical stability means the system's state converges to and stays within a neighborhood (a tube) around the target rather than reaching it exactly. On the robot, transverse coordinates converged to a neighborhood of the origin; its size is set by unmodeled contact forces, motor imperfections, vision noise, and delays.

Did the controller run on neuromorphic hardware?

No. The spiking law was computed in a conventional real-time loop: the Riccati solution and firing conditions were precomputed and B-spline interpolated, with state estimates from an encoder and a computer-vision system. The paper certifies a spiking control policy and validates it physically; it does not measure neuromorphic energy savings.

Why does this matter for organoid training?

It frames stimulation of living tissue as impulsive control of a hybrid system and gives a certificate-based firing rule: stimulate only when it provably decreases a chosen objective. The design pattern transfers even though the theorem's modeling assumptions do not; and the hardware result, that disturbances drive dense triggering, is a warning that closed-loop optogenetic protocols must budget for far more stimulation events than the event-driven ideal suggests.

References

  1. M. Surov and O. Sumenkov. Spiking Neuromorphic Control for Periodic Orbits Stabilization in Underactuated Robots: A Greedy Lyapunov-Based Approach. arXiv:2610.37310. 2026. https://arxiv.org/abs/2610.37310. Accessed 2026-10-31.