Back to search

Virginia Tech

Brain-Inspired Drone Navigation Using the Conflict Architecture

Abstract

dc:description.abstract

Novel forms of brain-inspired computer architectures are required both to understand the structural basis of intelligence and to overcome fundamental limits in computer parallelism and real-time autonomous decision-making. The brain does not compute a single answer and commit; it runs multiple, heterogeneous processing pathways in parallel on the same input, preserves conflicting candidate responses, and defers commitment until an external event or physical constraint forces selection. This organizational strategy is not arbitrary: as Rent's Rule [1-3] establishes for any large-scale system, the brain cannot sustain dense global synchronization without prohibitive communication costs, and so distributes computation across isolated, specialized modules — preserving local conflict rather than resolving it globally — until an external event forces commitment. Prior work introduced the Conflict Architecture (CA) to demonstrate that this principle can be instantiated in an engineered system, showing through maze-solving experiments that maintaining multiple conflicting algorithmic interpretations in parallel — and resolving conflict through time-driven arbitration rather than early convergence — can outperform any single strategy under unpredictable deadline constraints. That work established the CA as a proof of concept but left open the question of which classes of real-world problems benefit from this approach and whether the principle holds under a richer set of simultaneous physical constraints. This thesis addresses both questions. First, it systematically investigates the structural properties of problems for which the CA provides a principled advantage: critical decisions under unpredictable deadlines, multiple algorithmically distinct solution strategies, feasibility determined by physical constraints rather than purely on theoretical optimality, and continuous interaction with a dynamic and partially observable environment. This investigation motivates autonomous aerial navigation as a representative real-world domain, where unmanned aerial vehicles must simultaneously satisfy timing, energy, and safety constraints and where no single navigation strategy is optimal across all operating conditions. Second, this thesis introduces the CA-based Navigation Engine as a direct instantiation of the CA applied to UAV navigation in dynamic, obstacle-rich corridor environments. It runs three Algorithm Processing Elements (APEs) in parallel — a hard-failsafe reflex (APE1), a moderate planning depth reactive planner (APE2), and a high planning depth corridor navigator (APE3) — on a shared, time-stamped state snapshot, and commits whichever proposal is both complete and highest in planning depth before each externally imposed deadline expires. This organization instantiates Flynn's MISD[4] category — Multiple Instruction, Single Data — historically dismissed as either nonsensical because it produces conflicting answers or confined to fault- tolerant systems that treat those conflicts as errors to be caught. The CA reframes both: conflicting answers are not a defect but the design, and MISD becomes not merely coherent but advantageous precisely when the goal is to sustain multiple competing answers simultaneously until a physical deadline forces selection. The architecture is evaluated in a Gazebo–ROS 2 simulation of procedurally generated Manhattan-grid corridor worlds under a stochastic, event- driven deadline model that prevents any fixed strategy from reliably completing before action is required. Experiments across 1,000 simulation runs per strategy confirm and extend the core finding of the prior CA work. APE2 and APE3, deployed as standalone strategies, had event violations in over 80% and 86% of runs respectively — at mean violation rates of 28.02% and 40.39% respectively — rendering both non-feasible for deployment under real-time constraints despite their apparent speed advantages in uninterrupted conditions. Only APE1 and the CA achieved zero deadline event violations across all 1,000 runs. Between the two feasible strategies, the CA achieved a 34.55% reduction in mission elapsed time relative to APE1 (38.98 s versus 59.56 s; p < 0.0001) and a 34.52% reduction in algorithmic computation energy (97.51 J versus 148.92 J; p < 0.0001), with no statistically significant difference in no-fly zone violations or total physical flight energy. These results demonstrate that CA-style arbitration is Pareto- improving over the only other feasible standalone strategy. Significantly, the CA simultaneously reduces both completion time and computational energy cost, while conventional power reduction comes at the expense of reduced performance. This work extends the CA evaluation framework to include energy as an explicit metric alongside time and lays the foundation for further research in conflict-aware architectures applied to multi-constraint, real-time autonomous systems as well as a broader class of problems.

Degree

thesis:*
Name thesis:degree_name
Master of Science
Level thesis:degree_level
masters
Discipline thesis:degree_discipline
Computer Engineering
Department dc:contributor.department
Electrical and Computer Engineering
Grantor dc:publisher
Virginia Tech
Year dc:date.issued
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Chauhan, Vikhyat
Chair dc:contributor.committeechair
  • Paul, JoAnn Mary
Committee members dc:contributor.committeemember
  • Jones, Creed Farris
  • Tryfona, Nektaria

Subjects

dc:subject × 8

Rights

dc:rights
Statement dc:rights
  • In Copyright
Language dc:language.iso
en

Identifiers

dc:identifier.*
Dc Identifier Other
vt_gsexam:46571

Chain of custody

source
Harvested from
Virginia Tech
Base URL
vtechworks.lib.vt.edu/oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Chauhan, Vikhyat. Brain-Inspired Drone Navigation Using the Conflict Architecture. masters thesis, Virginia Tech, 2026. https://hdl.handle.net/10919/143136