Vel-v6.4: A Deterministic, Coordinate-Blind Multi-Agent Swarm Exploration Protocol for Constrained Hardware in GPS-Denied Topologies

Abstract

Autonomous multi-agent exploration in subterranean rubble, collapsed urban environments, and GPS-denied disaster topologies remains fundamentally constrained by the high resource demands of classical Simultaneous Localization and Mapping (SLAM) and centralized Coverage Path Planning (CPP). These traditional paradigms require megabytes of dynamic memory and dense global coordinate frames, making them unsuitable for deployment on low-cost, disposable micro-embedded edge hardware. This paper introduces Project Vel, a coordinate-blind, decentralized spatial exploration protocol operating under a strictly bounded O(1) time and spatial compute complexity. Project Vel completely eliminates global coordinate anchors and dynamic runtime memory allocation, executing within a 256 byte static .bss SRAM footprint. The protocol introduces three primary technical innovations: (1) Adaptive Gradient Clearance (AGC), a perception-driven local depth-sampling engine that overrides Kinetic Vector Flux (KVF) traps in non-rectangular and acute-angle geometries (< 90◦ ); (2) Dynamic Sparse Hash-Grid (DSHG), a zeroheap static slot pool utilizing floating-origin relative Morton keys and Least-Recently-Visited (LRV) ring eviction; and (3) Geographic Gossip Protocol (GGP), a delay-tolerant ad-hoc networking layer featuring monotonic state delta synchronization and densityadaptive broadcast throttling. Empirical validation across 500 Monte Carlo simulation runs demonstrates that Project Vel (v6.4) achieves a 100.0% escape success rate in acute rubble traps (compared to 5.0% in baseline models), eliminates runtime heap fragmentation crashes (0.0% failure rate), limits RF packet collisions to 2.1% under extreme node clustering, and restricts micro-aerial hover drift to 1.8 cm via active inertial counter-thrust braking.