Research analysis · Closed-loop control

Controlling a spiking system by its event timings alone

Every serious closed-loop biological-computing experiment has the same shape: read spikes, decide something, stimulate back. The hard half of that loop is not the reading; it is writing to excitable tissue in a way that respects its timing. Liu, Wroblowski, Alan, Shi, Sepulchre and De Schutter propose event-centric model predictive control, in which a mechanical scanning mirror driven by a two-neuron spiking oscillator is regulated using only the timestamps of beam-crossing events, with the neuronal parameters retuned at each event and the mechanical energy exchanged between actuator and mirror as the quantity being minimized. It is a numerical study on a non-biological plant, but it is the cleanest published formalization of what a stimulation loop around spiking hardware, or spiking tissue, would have to solve.

Source: Predictive Neuromorphic Control of a Scanning Mirror System: An Event-Centric Framework, arXiv:2611.11634v2, 8 November 2026. Primary source. Read: the full PDF, including Tables 1 to 3, the event-sensing construction, and the discussion of cost and model mismatch.

What the work claims

This is a primary methods paper from TU Delft, KU Leuven and Eindhoven University of Technology, extending the neuromorphic control line associated with Sepulchre's group.1 The claim has three parts. First, a control framework in which events define everything: two photodetectors placed almost exactly at the mirror's equilibrium position produce paired crossing detections, those paired events stand in for zero crossings and carry an estimate of angular velocity from their timestamp separation, the prediction horizon is a fixed number of future events rather than a fixed time window, and the control objective is phrased directly in event timings, prescribed zero-crossing times plus a desired swing amplitude. Event-triggered control in the existing literature only samples aperiodically; this framework makes the events themselves the state, the horizon and the target.

Second, the framework is integrated with a neuromorphic controller: a half-center oscillator of two spiking neurons, built on the mixed-feedback model of Ribar and Sepulchre, whose fast voltage crossing threshold directly gates the coil current of a reluctance-actuated scanning mirror. The tunable quantities are just three: the constant coil current level and two neuronal gain parameters that set spike frequency and spike width. Third, because no closed-form event-to-event prediction model exists for this coupled plant-plus-neuron system, the authors solve the optimization by sampling: an MPPI-inspired scheme draws 256 candidate parameter sequences over a horizon of 4 half-swings, refines over 2 rounds, rolls out the coupled ordinary-differential-equation model for each candidate, and applies only the first event interval's parameters before replanning at the next crossing. The paper explicitly notes that no certified optimality gap is established for this solver.

The result, in simulation: starting from the same untuned default parameters, the event-centric controller sustains oscillation at the reference 10 Hz with the prescribed zero-crossing times and a 5-degree scanning range while minimizing injected energy, and the nominal fixed-parameter controller cannot, eventually failing to supply enough energy to keep the mirror oscillating at all. The authors open-source their code.

How it works

The plant is a ferromagnetic mirror on a torsional flexure, a linear spring-damper oscillator with a natural frequency of 9.9 Hz and a damping ratio of 0.01, driven by reluctance torque proportional to current squared. The energy bookkeeping is the elegant part: the actuator supplies mechanical energy only when the mirror moves toward equilibrium and absorbs it when moving away, so sustaining oscillation at a prescribed amplitude means delivering precisely timed current pulses that top up the energy dissipated by damping. The two-neuron half-center oscillator generates those pulses sparsely: its fast voltage gates the current source, so actuation happens exactly at neuronal spikes and nowhere else. Between events the current level is held fixed, a deliberate concession to hardware in which a current source can only be adjusted intermittently.

Sensing is reduced to almost nothing. Two photodetectors sit at plus and minus gamma-s, where gamma-s is pi/3600, about 0.05 degrees of deflection, essentially on top of equilibrium. A zero-crossing event is declared when opposite detectors fire within 2.5 ms of each other; the crossing time is their midpoint and the angular velocity at crossing is estimated as 2 sigma gamma-s divided by their timestamp gap, where sigma carries the sign of the motion direction. No position encoder, no velocity sensor, no continuous waveform: the feedback channel is a handful of timestamps per second. The predictive layer then rolls out the coupled mirror-plus-neuron dynamics, whose neuronal time constants are 0.17, 1.75 and 17.46 ms, evaluates each sampled parameter sequence against event-timing error, amplitude error and mechanical energy transfer, and commits only until the next crossing, with a defined failure action of dropping the pulse amplitude to zero when no valid candidate is found, which the authors note does not guarantee sustained oscillation.

The evaluation is a single numerical example, implemented in Python 3.12 with adaptive Runge-Kutta rollouts: 50 inter-event intervals, about 2.5 seconds, a horizon of 4 half-swings equal to 0.2 s, nominal kinetic energy of the mover around 1.47 microjoules, and cost weights derived from that scale. Timestamps are assumed noiseless and sensing and computation delays negligible, assumptions the paper states openly.

Where a skeptic should push

The load-bearing assumption is that the optimizer can finish before the next event. The whole framework is real-time contingent: with a 50 ms inter-event interval and adaptive ODE rollouts of a stiff two-timescale system, the computation budget is generous here, and the authors concede that rollout cost limits the oscillation frequencies at which the approach can run, suggesting parallelism as the remedy. But the general lesson survives the example: event-centric control of fast spiking systems is a race between the prediction machinery and the tissue's own clock, and for biological event rates in the millisecond range that race is currently unwinnable with rollout-based optimization.

Second, this is simulation of a hand-built model, not hardware, and the same nominal parameters are used for plant and predictor, so no model mismatch is exercised at all; the discussion admits sensitivity to mismatch may accumulate over the prediction window and recommends shorter horizons without quantifying either. Third, safety is only partially handled: the angle constraint is enforced at event instants with no guarantee between them, and the failure action explicitly does not guarantee sustained oscillation, which for a mechanical mirror is a footnote and for a stimulation protocol would not be. Fourth, the comparison baseline is a deliberately untuned nominal controller; there is no comparison against a well-tuned classical controller or a learned one, so the paper demonstrates feasibility and value of online tuning, not superiority over strong alternatives. Fifth, the controller parameters fluctuate persistently without converging, which the authors attribute to the stochastic solver and to degenerate parameter combinations with equivalent performance; benign here, but a warning that event-centric loop signals may carry no stationary interpretation.

The write half of the closed loop

Organoid intelligence experiments that close the loop mostly close it in clock time: stimulate every 50 ms, or on a duty cycle, with patterns that ignore what the tissue is doing at the moment of write. The event-centric framework proposes the alternative contract that spiking substrate naturally suggests: define objectives in the timing of discrete events, sense only those events, and let the stimulation parameters update at the tissue's own event boundaries. An organoid expresses endogenous events that are directly analogous to the mirror's zero crossings, burst onsets, oscillatory phase zero crossings, population synchrony peaks. Recasting a training or steering objective as a desired sequence of such event times, with energy or dose as the minimized quantity, is a concrete, implementable research program, and none of its pieces requires solving the full state of the tissue, only a predictive model good enough to roll forward between events.

The deeper implication is architectural, and it cuts against a lazy narrative in the field. The tempting wetware story is that the living network is the controller and the surrounding machinery disappears. This paper shows what that machinery actually has to do: maintain a coupled model of plant and excitable circuit, roll it out hundreds of times per event, and run an optimizer under a hard real-time deadline. The neurons in this system are a sparse, energy-frugal actuation and timing layer; the intelligence of the loop sits in the digital predictor. That division of labor is, for now, the honest template for biological computing too: organoid tissue as the event-generating, energy-cheap dynamical core, and the digital layer as the predictive scaffold that writes to it. Claims that an organoid has learned to control something should be required to say which side of that line the learning happened on, because this architecture makes it entirely possible to demonstrate impressive closed-loop behavior in which the tissue contributed only excitability.

The threats deserve equal weight. Energy minimization as the stated objective has a dark twin in tissue: a controller rewarded for hitting event timings at minimal delivered energy will learn to steer an organoid toward whatever dynamical regime makes that cheap, which may be pathological hyperexcitability or quiescence, and the paper's own failure analysis, constraints guaranteed only at event instants, no guarantee of sustained safe operation, is a map of exactly where a stimulation loop can hurt a preparation between samples. The opportunity is a standards-level one: event timing, event-interval estimation from sparse sensors, dose accounting and a declared failure action are precisely the reporting elements missing from most closed-loop organoid papers, and adopting them would make results comparable across labs in a way that raw task performance never will.

The bottom line

Established, within a simulation: a scanning mirror with a two-neuron spiking controller can be regulated on frequency, amplitude and mechanical energy using only photodetector event timestamps, with the neuronal and current parameters retuned online by sampling-based predictive optimization over a 4-event horizon; the untuned nominal controller fails where the tuned one succeeds. Not established: hardware demonstration, superiority over well-tuned classical control, robustness to model mismatch or timestamp noise, or any path to millisecond-scale biological event rates with rollout-based solvers. The confirmation experiment for the organoid field is a stimulation loop that takes burst onsets or phase crossings as its events, updates stimulation parameters only at those boundaries, and reports event-timing error, delivered dose, and a declared inter-event safety policy. The breaking experiment is equally clear: show that event-centric updating beats well-tuned clocked stimulation on an organoid learning task at matched dose, or concede that the event machinery adds ceremony without capability.

Frequently asked questions

What makes this event-centric rather than just event-triggered?

In event-triggered control, events only decide when a conventional, trajectory-based controller updates. Here the events themselves are the feedback state, the prediction horizon is a count of future events, and the objectives are written in event timings, so sensing, prediction and regulation are all organized around the events rather than around clocked trajectories.

What is the neuromorphic part of the system?

The actuator timing layer: a half-center oscillator made of two spiking neurons whose fast voltage gates the mirror's coil current. Pulses are delivered exactly at spikes and nowhere else, giving sparse actuation, and the predictive controller tunes only three parameters, the current level and two neuronal gains that set spike frequency and width.

How can it predict without an event-to-event model?

By rolling out the coupled ordinary-differential-equation models of the mirror and the neurons themselves. A sampling-based, MPPI-inspired optimizer draws 256 candidate parameter sequences per event, scores each by simulated event timings, amplitude and energy, refines for 2 rounds, and applies only the first interval before replanning. The paper notes no certified optimality gap is established.

Is this demonstrated on hardware?

No. It is a numerical study in Python 3.12 with a single simulated example, noiseless timestamps, negligible sensing and computation delays, and identical plant and predictor parameters, so model mismatch is not exercised. The code is available on GitHub, and hardware validation is future work.

What would this look like for an organoid experiment?

Take burst onsets or oscillatory phase crossings as the events, define a training objective as desired event times, update stimulation parameters only at event boundaries, and report event-timing error, delivered dose, and an explicit safety action for intervals between events. The mirror study is essentially a specification of the write side that such a loop would need.

What is the main limitation for biological use?

Computation must finish before the next event. With rollout-based optimization, the approach suits the mirror's 50 ms event spacing; millisecond-scale biological events are out of reach for that solver class today. Short-term plasticity, noise, and drift in the tissue would also punish the model-mismatch assumptions this study leaves untested.

References

  1. C Liu, R Wroblowski, A Alan, S Shi, R Sepulchre, B De Schutter. Predictive Neuromorphic Control of a Scanning Mirror System: An Event-Centric Framework. arXiv:2611.11634. 2026. http://arxiv.org/abs/2611.11634v2. Accessed 2026-11-11.