IEEE RAEEUCCI 2026 · ★ BEST PAPER

SWARM COORD

Event-triggered decentralized swarm coordination for multi-vehicle collision avoidance — communicate only when it matters, and still hit zero collisions.

Sreedharshan G J, Surya Narayanan S, Akankshya Sethi, Aravindan M · SRM Institute of Science and Technology

IEEE BEST PAPERMULTI-AGENT PSO-DMPCSUMO SIMULATION2026

Paper / DOI ↗GitHub — not publicly sharedPoster — not publicly shared

01 Problem

Decentralized multi-vehicle coordination usually assumes constant, all-to-all V2V communication between autonomous vehicles (AVs) to avoid collisions. That doesn't hold up under real bandwidth constraints: continuous onboard optimisation and constant broadcasting place a growing strain on both communication bandwidth and compute as fleet size increases.

02 Motivation

Existing event-driven platoon controllers react to state-deviation thresholds but don't tie the trigger to any solver-initialisation strategy, leaving per-step compute cost fixed. Prior PSO-DMPC hybrids optimize every time step regardless of traffic conditions, so overhead stays constant irrespective of how often a genuine safety risk actually arises. This work asks whether the scarcity of near-collision events can be exploited directly: trigger the AI layer only when a safety margin is actually violated, and use the PSO solution to warm-start the local DMPC solver rather than solving from scratch.

03 Research Question

Can an event-triggered mechanism decide, locally and per-agent, when a communication update is actually necessary for safe decentralized coordination — cutting communication and compute overhead without sacrificing collision-avoidance guarantees?

04 Architecture

A decentralized multi-agent system combining Particle Swarm Optimisation (PSO) with Distributed Model Predictive Control (PSO-DMPC). Each vehicle node continuously estimates its own state (x, y, v, θ) under the kinematic bicycle model and checks the inter-vehicle distance against a delay-compensated safety margin. A PSO routine — with a five-step anti-chattering cooldown — is triggered only when that margin is violated, providing a warm-start to the local DMPC optimizer rather than running continuously.

System architecture diagram of the AI-based event-triggered PSO-DMPC swarm framework, showing the SUMO simulation environment, per-vehicle state estimation, safe-distance violation check, PSO speed optimizer, DMPC controller, and IEEE 802.11p DSRC V2V communication path.

Fig. 1 — System architecture: each vehicle node runs state estimation → safe-distance check → (if violated) PSO speed optimizer → DMPC controller, communicating over IEEE 802.11p DSRC (5.9 GHz, 300 m range) only when a trigger fires. Activation rate 7.4% · channel utilisation 5.2%. Reproduced from the accepted manuscript.

Event-Triggered PSO Speed Optimisation (Algorithm 1) — for each triggered agent:

Initialise a swarm of N_p = 5 particles, velocities drawn uniformly from [v_min, v_max]; best fitness ← ∞.
For 5 iterations, evaluate the fitness function (Eq. 4 below) at each particle and track the personal/global best.
Return the best candidate speed as the DMPC warm-start.
Bounded by a 50 ms computation budget per trigger, making the routine embedded-hardware feasible.

The PSO subsystem only supplies a warm-start guess — it adds no additional hard constraints — so it cannot disturb the closed-loop stability properties already guaranteed by the DMPC layer.

05 Figures

Bar chart of speed variance for Baseline, Continuous PSO, and Event-Triggered PSO, alongside a pairwise statistical-significance matrix of p-values.

Fig. 3 — Speed variance: three-way comparison (baseline 17.58, Continuous PSO 3.20, ET-PSO 3.19 m²/s²) with the pairwise significance matrix. Both AI variants beat baseline at p<0.001; ET-PSO vs. Continuous PSO is not significantly different (p=0.101), meaning the 84% compute saving comes at no measurable cost in speed stability.

Bar chart showing Mean Time to Collision safety-margin improvement over baseline for Continuous PSO (+82.7%) and Event-Triggered PSO (+90.5%).

Fig. 6 — MTTC safety-margin improvement vs. baseline: Continuous PSO +82.7% (37.0 s), Event-Triggered PSO +90.5% (38.6 s) — both p<0.001.

Bar chart comparing AI activation rate and transmission load as a percentage of continuous PSO, for Baseline, Continuous PSO, and Event-Triggered PSO.

Fig. 11 — Communication efficiency vs. Continuous PSO: 7.4% AI activation rate (Event-Triggered PSO) achieves only 88.0% of Continuous PSO's transmission load — confirming the 92.5% reduction in active optimisation steps. Figures reproduced from the accepted manuscript.

06 Equations

$$d_{safe} = \max\left(d_{min},\ v_i\cdot(\tau_{reaction}+\tau_{comm}+\tau_{PSO})\right)$$

Delay-compensated safety margin. At v_max=18 m/s, τ_reaction=1.5 s, τ_comm=20 ms, τ_PSO=50 ms: d_safe=28.26 m, with maximum tolerable latency τ_max=670 ms.

$$\min_{u_i} \sum_{k=0}^{N-1}\Big(\lVert x_i(k)-x_{ref,i}(k)\rVert_Q^2+\lVert u_i(k)\rVert_R^2\Big)+\lVert x_i(N)-x_{ref,i}(N)\rVert_P^2$$

DMPC objective, subject to vehicle dynamics, the inter-agent separation constraint ‖x_i−x_j‖ ≥ d_safe for all j≠i, and input bounds u_min ≤ u_i ≤ u_max.

$$f(v_c) = 0.6\cdot\max(0,\, d_{safe}-d_{ij}) + 0.4\cdot|v_c - v_i|$$

PSO fitness function for the triggered speed search — balances closing the safety-margin gap against deviating from the current speed.

$$V_i(x_i(k{+}1)) - V_i(x_i(k)) \le -\lVert x_i(k)-x_{ref,i}(k)\rVert_Q^2 \le 0$$

Lyapunov non-increase condition for the DMPC closed loop, giving asymptotic stability of the tracking error between triggered updates.

07 Experimental Setup

Validated in SUMO (Simulation of Urban Mobility) over a 2×2 orthogonal grid network, 500 simulation steps at Δt = 0.1 s per run, across 20 independent random seeds — over 10,000+ simulation timesteps in total. Three configurations were compared: Baseline (SUMO car-following without AI), Continuous PSO (optimizes every step), and the proposed Event-Triggered PSO.

Default parameters: free-flow speed v ∈ [3, 18] m/s, minimum separation d_min = 10 m, N = 10 vehicles, N_p = 5 particles, I = 5 PSO iterations, 5-step anti-chattering cooldown. V2V communication modeled over IEEE 802.11p DSRC (5.9 GHz, 300 m range, 20 ms latency).

08 Results

Metric	Baseline	Continuous PSO	Event-Triggered PSO
Speed variance (m²/s²)	17.58 ± 0.04	3.20 ± 0.01	3.19 ± 0.01
MTTC (s)	20.25	37.00 (+82.7%)	38.57 (+90.5%)
CPU cost / step	0.12 ms	2.80 ms	0.45 ms (−84%)
Near-miss events	3083	2800	2748 (−10.9%)
Collisions (10,000+ timesteps)	—	—	0

ET-PSO matches Continuous PSO's speed-variance and safety performance (p=0.101, not significant) at 84% lower per-step compute cost. All differences vs. Baseline are significant at p<0.001. Formation error (target 15 m spacing) drops 48.4% under ET-PSO (62.5 → 32.3 m) versus Baseline, though full string stability (disturbance amplification <1) was not achieved in every run.

Note from the manuscript: SUMO's default collision-handling protocol does not simulate physical impacts — it teleports or removes vehicles instead — so the reported zero-collision result reflects the absence of safety-constraint violations (d_ij < d_min), not literal impact absence.

09 Ablations & Sensitivity

PSO parameter sensitivity (6 configurations, 20 trials each, 120 trials total): speed variance ranges 2.31–2.64 m²/s² across inertia/cognitive/social-weight variants — all within one standard deviation of the default (2.40 ± 1.28) — with zero collisions in every configuration.

Config	w	c₁	c₂	Variance (m²/s²)
Default	0.7	2.0	2.0	2.40 ± 1.28
Low inertia	0.4	2.0	2.0	2.64 ± 1.24
High inertia	0.9	2.0	2.0	2.35 ± 1.38
High social	0.7	2.0	2.8	2.31 ± 1.21

Density robustness (5–26 vehicles, 20 trials each): speed variance grows sub-linearly with fleet size (1.59 → 10.70 m²/s²), with zero collisions at every density tested — evidence of graceful degradation rather than a hard failure point.

10 Failure Cases

At the highest tested density (26 vehicles), speed variance jumps to 10.70 m²/s² — noticeably higher than the 1.59–3.19 m²/s² range seen at lower densities, the roughest regime the framework was tested against. Full string stability (disturbance amplification factor < 1) was not obtained in every experimental run, even though overall group cohesion improved substantially over baseline. And because SUMO's collision handling teleports vehicles rather than modeling physical impact, the "zero collisions" result should be read as zero safety-constraint violations rather than a literal crash-test guarantee.

11 Limitations

String stability was not achieved with speed-only control in every scenario.
Formal Lyapunov/Control-Barrier-Function proofs are deferred to future work — the current safety argument is a practical, empirically-supported one rather than a closed formal proof.
PSO weights (inertia, cognitive, social) are manually tuned, not learned.
Cyber-attack and denial-of-service resilience of the triggered communication channel is not modelled.
Validation is simulation-only (SUMO), with embedded-hardware feasibility (NVIDIA Jetson) argued from the 50 ms compute budget rather than demonstrated on physical hardware.

12 Future Work

A density-adaptive safety threshold d_safe(ρ, v) that responds to local traffic density rather than a fixed margin.
Formal Lyapunov/CBF stability proofs to close the gap left by the current practical-invariance argument.
Hardware-in-the-loop validation on embedded platforms (e.g. NVIDIA Jetson) to confirm the simulated 50 ms compute budget holds on real hardware.
Modelling denial-of-service and cyber-attack resilience for the event-triggered DSRC channel.