Computation as a controlled loss of indecision
A decentralized controller lets groups of robots resolve traffic conflicts by holding them near a state of collective indecision until the situation becomes safety-critical, then letting a bifurcation decide who goes first. Tested in 250 simulated rollouts with zero safety failures and on six physical robots, it also reads as a design pattern for computing with organoid neural tissue.
Source: Neuromorphic Decision Dynamics for Deadlock-free Safe Navigation in Congested Environments, arXiv:2610.33271 [cs.RO], October 23, 2026. Primary source. Read the full arXiv HTML version including supplementary text.
What the work claims
This is a control-theory paper with proofs, large simulations, and hardware validation. Sinhmar, Srivastava, and Leonard introduce what they call neuromorphic decision dynamics (NDD): a decentralized, event-based controller in which each agent in a congested environment carries an internal decision state and an attention state, integrates locally observed relative motion into a continuous measure of conflict evidence, and modulates its speed along its prescribed path. No agent communicates, no priorities are preassigned, and no agent models its neighbors' intentions1.
The central claim is that safe and live deconfliction can be made an emergent, tunable property of the local interaction dynamics rather than a planned solution. When a group's conflict is mild, the agents' decision states all sit near zero, a symmetric equilibrium the authors call collective indecision. As conflict pressure rises past a threshold set by design parameters, that equilibrium loses stability through a bifurcation, and differentiated decision states emerge: some agents' internal states grow positive, others negative, and the corresponding speed changes separate the agents' projected entry times into the contested region. A traversal order forms, in finite time, with guarantees of no deadlock and no livelock for finite heterogeneous populations of arbitrary changing size. Conditions on sensing range, feedback gains, response time scales, and actuation limits that make this work are derived explicitly; the supplementary text lists five compatibility conditions on controller parameters, a choice the authors note is independent of population size1.
The evidence has three tiers. In simulation against baselines (optimal reciprocal collision avoidance, LivePoint, and centralized and decentralized control-barrier-function quadratic programs), 50 randomized start-goal scenarios per population size at N in {4, 8, 12, 24, 32} give 250 NDD rollouts with a 100 percent goal-reaching success rate and pairwise separation always above the safety threshold; every baseline had lower success, and ORCA and both CBF-QP variants showed mean minimum separation below the safety threshold. Two deliberately brutal 15-agent scenarios, a near-antipodal exchange and nearly parallel intersecting groups, produced collisions in all baseline methods and none in NDD. On hardware, six TurtleBot3 robots with a Vicon motion-capture feed and no inter-robot communication coordinated crossings at a nominal 0.2 meters per second, and two antipodal trials whose starting positions differed by only 0.8 to 2.5 centimeters per robot produced different traversal orders, showing the order genuinely emerges rather than being scripted.
How it works
The mechanism has three ingredients worth separating. First, evidence integration: each agent converts the observed relative motion of nearby agents into a scalar conflict-pressure signal, weighted across its simultaneous interactions so that its most safety-critical conflict dominates. Second, decision-attention dynamics: the internal decision state rests near zero while pressure is low and departs rapidly once pressure crosses the bifurcation threshold, with an attention state gating how strongly evidence drives the decision. Third, decision-action feedback: the decision modulates speed, the speed change alters relative motion, and the altered motion changes the evidence every agent sees. It is this closed loop, not any single agent's computation, that creates the temporal separation resolving the conflict. Once agents clear the contested region, evidence decays, decisions relax back toward zero, and nominal motion resumes.
Two design ideas do the heavy lifting. Group formation is implicit: agents sharing conflicts self-organize into disjoint locally engaged groups that resolve concurrently, so global complexity never enters any single computation. And deadlock is handled not as a failure to detect but as a state to make unstable: near-symmetric configurations sit near collective indecision, and the bifurcation amplifies infinitesimal asymmetries in motion cues into a rapid, decisive differentiation. The hardware trials make the same point physically: 0.8 to 2.5 centimeters of initial-position difference was enough to tip which robot passed first, twice, differently.
Where a skeptic should push
The most load-bearing assumption is that agents can faithfully estimate the relative motion that drives conflict evidence. On the real robots this estimation came from a motion-capture system with globally accurate positions and low-pass-filtered finite differences for velocity: excellent sensing that a deployment scenario will not have. The proofs accommodate bounded sensing noise and actuation error, but the hardware demonstration is a single six-robot indoor setup on prescribed straight routes, one lab, one platform family. That is a proof of deployability, not of robustness in the field.
Second, the word neuromorphic should be read generously. No spiking hardware appears in the paper; the neuromorphic claim is by analogy to decision-making dynamics and event-driven local computation. A purist could reasonably file this under nonlinear decentralized control, and the burden is on the authors' framing that the dynamical perspective is the substantive connection. Third, the bifurcation's sensitivity is a feature for traffic and a liability elsewhere: when 2.5 centimeters decides who goes first, the output is path-dependent and effectively stochastic at the resolution any downstream system can measure. Safe passage tolerates that; a computation that must be repeatable does not, at least not without an external mechanism for holding initial conditions.
Finally, the guarantees, while explicit, are conditional: they hold when the five compatibility conditions on gains, sensing range, response times, and actuation limits are jointly satisfied, and tuning to them requires platform characterization. The paper provides a practical tuning workflow in its supplement, but certification effort is real, and heterogeneity across agents enters the conditions nontrivially.
Symmetry breaking as computation in neural tissue
The connection to organoid intelligence is not the robots; it is the computational pattern, which is genuinely neuromorphic in a deeper sense than the paper's own usage. NDD computes by parking a dynamical system at a symmetric unstable equilibrium, letting input evidence determine when and how the symmetry breaks, and reading the answer off from which attractor wins and how the population differentiates. That is a fair one-paragraph description of what populations of biological neurons are known to do in decision tasks: accumulated evidence destabilizes a balanced state, and the network falls to one side. The paper contributes something the neuroscience account usually lacks: a constructive recipe. It specifies how to choose the feedback gains, the time-scale separation between evidence accumulation and commitment, and the gating that keeps weak evidence from triggering premature decisions, and it proves which combinations of those choices guarantee the symmetric state is destabilized exactly when intended1.
For an organoid on a closed-loop electrode array, this suggests a concrete and largely unexplored programming model. Rather than decoding spike trains into symbols, an operator could bias the tissue's collective state with patterned stimulation, hold it near a critical balanced state, and read decisions from coarse population activity: which assemblies amplify, which suppress, when. The readout problem, the hardest part of organoid computing, becomes easier under this model because bifurcation timing and winner identity are population-level observables, robust to single-electrode noise in a way that precise spike-timing codes are not. The paper's grouping mechanism has a biological echo too: interacting agents self-organize into disjoint engaged groups, much the way neural populations form transient functional assemblies, suggesting that multi-problem parallelism could be an emergent property rather than a scheduling problem.
The threat side is the certificate asymmetry, and it is sharp here. NDD's behavior is guaranteed because its parameters can be tuned to satisfy five explicit compatibility conditions chosen independently of population size. Living tissue offers no such tuning interface: you cannot set a cortical organoid's effective feedback gain, and its time-scale separation is set by development and physiology, not by design. An organoid might exhibit beautiful bifurcation-based decision dynamics while offering no way to know, from first principles, when indecision will hold and when it will break. That is a real research program, not a dismissal: the paper's conditions could be inverted into measurement targets, asking which stimulation protocols move a culture's effective parameters toward the certified regime, and using the bifurcation signatures (sensitivity to microscopic asymmetry, relaxation back to baseline after resolution) as assays of computational state.
The dual-use note belongs in any honest reading. A mechanism whose purpose is to destabilize collective indecision in a crowd, with tunable sensitivity to tiny cues, is also a mechanism for understanding how consensus gets manufactured or disrupted in any population of coupled decision-makers, biological or human. Nothing in the paper touches that, but the mathematics does not care what the agents are.
The bottom line
As robotics, this is a strong paper: a constructive, provable answer to making deadlock an unstable state rather than a detected failure, validated from 32-agent simulations down to centimeter-scale sensitivity on physical robots. Its limits are honest ones: idealized sensing, straight prescribed routes, six robots, and conditional guarantees that demand real tuning effort. For organoid intelligence, its lasting value is a transferable abstraction: computation as controlled symmetry breaking, with a specification for the feedback, gating, and time-scale parameters that make it reliable. Whether a culture of living neurons can be steered into, and characterized within, that regime is an open experimental question that current closed-loop MEA setups are, for the first time, equipped to ask.
Frequently asked questions
What are neuromorphic decision dynamics?
A control method in which each agent integrates locally observed motion into a conflict-evidence signal that drives an internal decision state. When conflict pressure crosses a threshold, the group's shared state of collective indecision loses stability through a bifurcation, and differentiated decisions emerge that sequence the agents' passage. Agents never communicate or receive preassigned priorities.
How well did it perform?
Across 250 simulated rollouts at group sizes from 4 to 32 agents, every NDD agent reached its goal with pairwise separation always above the safety threshold, while all compared baselines (ORCA, LivePoint, centralized and decentralized CBF-QP) had lower success, with ORCA and both CBF-QP variants falling below the safety threshold on mean minimum separation. Hardware tests with six TurtleBot3 robots showed coordinated crossings without communication.
What was the most striking hardware result?
Two antipodal-crossing trials whose robot starting positions differed by only 0.8 to 2.5 centimeters produced opposite traversal orders, with the same robot passing first in one trial and last in the other. This shows the passage order is genuinely emergent: microscopic asymmetries in motion cues are amplified by the decision-action feedback loop.
How can deadlocks be avoided rather than detected?
By design, the controller makes the deadlocked configuration an unstable equilibrium. Near-symmetric stalemates sit close to collective indecision, and the bifurcation amplifies tiny motion asymmetries into decisive differentiation, so unresolved configurations destabilize themselves as agents move. The paper derives sensing, gain, and actuation conditions that guarantee finite-time resolution without deadlock or livelock.
Does this use spiking neural networks or spiking hardware?
No. The neuromorphic label refers to the dynamical style of computation (event-driven, evidence-accumulating, bifurcation-based), not to spiking neurons or neuromorphic chips. The controllers run as ordinary continuous-time dynamical systems on each robot.
Why does this matter for computing on living neural tissue?
It provides a constructive blueprint for a style of computation that living neural tissue is already suspected to perform: hold a network near a balanced state, let evidence control when symmetry breaks, and read the answer from population-level differentiation. That reframes the organoid readout problem from decoding precise spike timing to monitoring coarse population states, and it specifies measurable parameters (feedback strength, time-scale separation, gating) that closed-loop electrophysiology could target.
References
- H. Sinhmar, V. Srivastava, N. E. Leonard. Neuromorphic Decision Dynamics for Deadlock-free Safe Navigation in Congested Environments. arXiv:2610.33271 [cs.RO]. 2026. https://arxiv.org/abs/2610.33271. Accessed 2026-11-08.