One numerical test of a multi-agent control method reported a 4.6-fold reduction in the number of safety constraints retained. The reduced formulation kept an average of 2.80 constraints, compared with 13 in the full formulation. The result came from one closed-loop numerical scenario.
The same comparison reported identical values for three measures: a final formation error of 0.070722 metres, minimum obstacle clearance of 0.2721 metres and minimum distance between agents of 1.75 metres. The paper presents the work as a methods study. The supplied document is arXiv version 1, dated 26 August 2026.
How the reduction is certified
The paper targets tube-tightened multi-agent eCBF safety halfspaces. In ordinary terms, it seeks to keep a smaller set of safety inequalities while showing that the controls allowed by that smaller set remain inside the controls allowed by the full tightened set. The central idea is a certified, geometry-adaptive reduction.
Geometry determines the starting point. For planar inputs, the largest-angular-gap rule retains two extreme directions in the strict half-plane regime. In other geometries, the construction starts with three retained constraints and adds more if the certification check does not pass.
That certification check uses a Farkas certificate, a mathematical test of containment. It verifies that every control admitted by the reduced set is also admitted by the full tightened set. The associated conic multipliers and nominal-aware offsets are obtained in closed form, without an auxiliary optimization.
The safety statement is conditional. Under the stated modeling assumptions, the tube-validity assumption and the required eCBF initial conditions, the theorem states that satisfying the tightened constraints makes the original safety sets forward invariant for the true disturbed system.
The numerical comparison
The numerical comparison involved ten follower agents and four obstacles. The model used a follower jerk-servo model with five desired-acceleration fields. The reduced formulation was assessed against the full tube-tightened formulation in a closed-loop scenario.
In that scenario, the full and reduced formulations reported the same final formation error, minimum obstacle clearance and minimum inter-agent distance. The values were 0.070722 metres, 0.2721 metres and 1.75 metres, respectively.
The reported nominal-control rejection rate was 2.5% for the reduced formulation and 2.7% for the full formulation. Reported constraint violations were 0.0% for both.
Alongside the numerical comparison, the paper gives a separate result about nominal controls. Its corollary states that if a nominal control is admissible for the full tightened set, it remains admissible for the reduced set. The implementation description says retained constraint indices can be reused when possible, current-stage geometry quantities are recomputed, and the exact full-set formulation can be restored.
Synthetic timing sweeps showed a wider gap
A separate set of synthetic timing sweeps examined solve-time growth as the constraint count and prediction horizon varied. In one sweep, the reported growth was about 147-fold for the full formulation versus 4.7-fold for the reduced formulation. In the other, it was about 458-fold versus 7.0-fold.
Some reduced-formulation instances were escalated, with a reported rate of 2.4%. In the method, escalation means adding retained constraints when the initial certification check fails. The procedure can ultimately restore the exact full set.
The result has a narrow reach
The numerical evidence is limited to one ten-follower, four-obstacle closed-loop scenario and synthetic timing sweeps. The supplied analysis does not report the number of runs or variability for the closed-loop reduction or the timing results.
The construction described here is for planar inputs. The supplied evidence does not establish performance beyond that setting, the reported numerical scenario or the synthetic timing sweeps. It also does not establish that the reduced formulation will always remain feasible, since the design includes enlargement and full-set fallback when needed.
Funding information and a conflict-of-interest statement are not reported in the supplied document.
Paper data and sources
Original title: Scalable Tube-Tightened Multi-Agent Safety via Certified Constraint Reduction
Authors: Armel Koulong
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text