active Simulation and planning · physical flights later · 11 log entries

Part of ggSkyGrid · See the connected roadmap →

GG Swarm

Decentralized drone-swarm research, moving from a simulated RL capstone toward a real swarm built on a certifiable planning core.

GG Swarm

Objective

Build a legitimate decentralized swarm: a deterministic plan-and-certify core first, then online replanning behind a safety filter, with the learned GATv2/PPO policy benchmarked as a challenger.

Current Phase: Simulation and Planning (Post-Capstone)

After the capstone I asked how real drone swarms get built. Three independent surveys agreed: swarms that work in the field run on predictable, checkable planning code, not end-to-end learning. So the plan changed. A deterministic plan-and-certify core comes first. It assigns drones to formation slots, plans the transitions and proves them safe before anything flies. Online decentralized replanning comes next, behind a safety filter, and the GATv21/PPO2 policy from the capstone runs behind that filter as a challenger, benchmarked against the classical methods. Hardware goes Crazyflie indoors first, then an ArduPilot outdoor drone. Simulation and planning are active. The rover remains the main bench build, with physical swarm work following the single-aircraft and simulation milestones.

Cinematic Trailer

Cut from twenty clips captured in Phase 5 of the capstone: decentralized GATv2/PPO policies holding formation, avoiding obstacles and coordinating in simulation.

Roadmap & Development Phases

Active · 6 Milestones · Plan-and-Certify First
Phase 0 Complete

Capstone Baseline

v1.0.0-capstone: GATv2/PPO formation policy in Isaac Lab.

Phase 1 Complete

Plan-and-Certify Core v1

Built in ggChoreo: slot assignment, smooth transitions and a validator that proves separation, speed limits and geofence before flight.

Phase 2 Planned

CrazySim Test

Fly certified plans in the Crazyflie simulator and compare flown against planned.

Phase 3 Planned

Crazyflie Indoors

One Crazyflie first, with repeatable missions. Add aircraft only after multi-drone simulation and positioning requirements are demonstrated.

Phase 4 Planned

Decentralized Replanning

Drones react to each other live behind a safety filter; the RL policy is benchmarked head to head.

Phase 5 Stretch

Outdoor Swarm

ArduPilot outdoor drones, shared with the ggSkylight show platform.

Archived Milestone: Academic Capstone (v1.0.0)

This section preserves the original research and simulation work completed for my Computer Science Capstone at CSUMB.

Project Overview

The capstone addressed a critical bottleneck in the deployment of unmanned aerial vehicle swarms by tackling the inherent vulnerabilities found in centralized control architectures. The project proposed a fully decentralized coordination framework where global formation behavior emerges naturally from local agent interactions.

Technical Architecture

The architecture was split into the Brain (GATv2 spatial reasoning) and the Muscles (MINCO3 trajectory optimization), unified by a GNSC 5-Layer model.

GNSC 5-Layer Architecture

flowchart BT
    L1["<b>Layer 1: Local Sensing</b><br/>12D body-frame + K×3 neighbor rel_pos"]
    L2["<b>Layer 2: GNN Message Passing</b><br/>2-layer GATv2, K=2 sparse edges, edge cache"]
    L3["<b>Layer 3: Distributed Consensus</b><br/>MINCO min-jerk filter (T=0.04s) + SwarmRaft dropout"]
    L4["<b>Layer 4: Runtime Safety Shields</b><br/>CBF barrier constraints, clamped corrections, MINCO sync"]
    L5["<b>Layer 5: Mission Execution</b><br/>Thrust/moment mapping → physics"]

    L1 --> L2 --> L3 --> L4 --> L5

    style L1 fill:#3498db,color:#fff
    style L2 fill:#2ecc71,color:#fff
    style L3 fill:#f39c12,color:#fff
    style L4 fill:#e74c3c,color:#fff
    style L5 fill:#8e44ad,color:#fff

Capstone Timeline & Simulation Performance

The simulation phase leveraged NVIDIA Isaac Lab4 for GPU-accelerated physics, achieving high-fidelity results in formation stability and obstacle avoidance.

  • Mean Formation Error: < 0.1m during steady flight.
  • Success Rate: > 95% across randomized obstacle-dense environments.

Capstone Timeline

  • Weeks 1–4: Proposal and Plan (Completed)
  • Weeks 5–11: Core Development (Brain/Muscles) (Completed)
  • Weeks 12–15: Stress Testing and Showcase Prep (Completed)
  • Week 16: Capstone Festival Delivery (Completed April 2026)

References

  1. Brody, Alon, Yahav (2022). “How Attentive Are Graph Attention Networks?”. International Conference on Learning Representations (ICLR). arXiv:2105.14491 ↑

  2. Schulman, Wolski, Dhariwal, Radford, Klimov (2017). “Proximal Policy Optimization Algorithms”. arXiv:1707.06347 ↑

  3. Wang, Zhou, Xu, Gao (2022). “Geometrically Constrained Trajectory Optimization for Multicopters”. IEEE Transactions on Robotics 38(5): 3259–3278. doi:10.1109/TRO.2022.3160022 ↑

  4. Mittal, Yu, Yu, Liu, et al. (2023). “Orbit: A Unified Simulation Framework for Interactive Robot Learning Environments”. IEEE Robotics and Automation Letters 8(6): 3740–3747. The framework that became NVIDIA Isaac Lab. doi:10.1109/LRA.2023.3270034 ↑

Project Log

11 entries
← All projects