KERES — Multi-Drone Simulation & Experimentation
I helped turn a broad swarm concept into a configurable simulation platform engineers could test, measure, and replay.
- Contribution
- Coordination, PX4 control, experiment infrastructure
- Team
- Booz Allen systems engineering internship
- Constraints
- Simulation prototype; protected operational context
- Outcome
- 7 strategies, 6 formations, 417 passing tests
Current phase: formation.
Scope and boundary
A research platform, not a deployed system.
The public case study separates implemented software from long-term concepts and unvalidated runtime paths.
What the prototype investigated
- Simulated multi-drone coordination and defensive formations
- PX4-backed and lightweight simulation control paths
- Repeatable scenarios, structured events, metrics, and replay
- Simulator-independent coordination and validation modules
What this page does not claim
- Physical-drone deployment or validated operational performance
- Weapons, damage, lethality, or real threat intelligence
- A production autonomy stack or production Cesium integration
- A claim that every supported configuration ran in the full GPU simulator
What I worked on
I worked across the system boundaries.
Swarm coordination
Implemented and integrated formation, assignment, tracking, interception, and matched-seed benchmarking logic across several coordination approaches.
Simulation & PX4 control
Connected swarm-level commands to simulated multirotors through Pegasus and PX4 SITL, including OFFBOARD lifecycle management and coordinate conversion.
Experiment infrastructure
Built versioned run records, configuration hashing, run bundles, event recording, integrity checks, and browser-based replay for reproducible experiments.
Simulation environments
Worked on USD environment ingestion, obstacle handling, and visualization seams while keeping environment dependencies isolated from core simulation behavior.
System architecture
Simulation, coordination, control, and evidence stay connected.
Select a stage to see its responsibility. Teal marks the software path Brian helped build, integrate, or harden.
Assigns simulated interceptors and produces high-level position, velocity, and yaw intent.
Engineering scale
The 126 figure describes implemented coordination × formation × threat-profile configuration space before seeds, swarm sizes, or flight backends. It is not a claim that every combination was validated in the full Isaac, Pegasus, and PX4 runtime.
Formation explorer
Compare the geometry without turning it into spectacle.
Formation families were interchangeable experiment inputs. The useful question was how geometry changed coordination behavior under controlled scenarios.
The explorer shows implemented formation families schematically. It does not represent measured operational coverage.
Technical deep dives
Tactical intent stays separate from flight control
KERES produced position, velocity, and yaw commands while PX4 owned estimation and low-level vehicle control. The PX4 path covered OFFBOARD pre-streaming, mode switching, arming, heartbeat monitoring, diagnostics, forced disarming, and cleanup.
A run becomes inspectable evidence
Scenario identity, configuration hashes, seeds, tool versions, events, time series, and integrity-checked artifacts made runs reproducible. Episode JSONL could be converted into a self-contained browser replay.
Event order protects the result
Safety and breach accounting occurred before interception resolution, preventing a hostile that had already reached the asset from being counted as a successful interception.
Verification approach
Most logic stayed testable without the GPU simulator.
- 01417 local tests passed on the merged integration tree.
- 02Ruff linting, formatting, mypy, and JSON Schema validation were included in the recorded verification.
- 03Core coordination, geometry, recording, safety, and validation logic remained testable without launching the GPU simulator.
- 04Four coordinated-threat characterization cases reproduced archived control mathematics with a maximum absolute difference of 0.0.
Result
From concept to measurable experiments.
KERES finished the internship as a simulation prototype with a broader software foundation than the initial concept. It could compare several coordination strategies and formations, execute high-level commands through PX4-backed simulated vehicles, record structured experiment evidence, and replay mission behavior for analysis.
My work focused on the software layers that made those experiments configurable, testable, and reproducible. The project is the clearest example in my portfolio of working across simulation, controls, multi-agent logic, observability, and verification.