ADR-148: Drone Swarm Control System — Topologies, Strategy Formulations, Self-Learning & Vertical Applications

Scope note: ADR-147 deferred Cosmos WFM to "ADR-148" as an offline data generator. That item is promoted to ADR-171 (the swarm-benchmarking/evaluation companion to this ADR; renumbered from ADR-149 to resolve the ADR-149 duplicate-number collision). This ADR takes 148 to address the broader drone swarm control architecture, which is the first consumer of ADR-147's OccWorld occupancy output.

1. Context

1.1 Motivation

ADR-147 established a validated 3D occupancy world model (OccWorld, 1.65 GB VRAM, 375 ms/inference) that predicts future-state voxel occupancy from WiFi CSI. That output — a spatiotemporal occupancy grid at 0.2 m/voxel — contains the environmental and human state information required to plan drone swarm missions. No architecture currently bridges ADR-147's world model to airborne agents.

The wifi-densepose-signal pipeline (ADR-134 CSI→CIR, ADR-135 calibration, ADR-146 RF encoder) already achieves real-time human detection via ESP32-S3 + companion compute. The next logical extension is deploying this sensing stack as an airborne payload across a coordinated drone swarm, enabling:

• Search-and-rescue (SAR) localization through debris and walls
• Precision area coverage that adapts in real time to detections
• Persistent environmental monitoring without fixed infrastructure

No existing ADR covers drone fleet coordination, swarm topologies, MARL-based autonomy, or the regulatory compliance requirements for beyond-visual-line-of-sight (BVLOS) operations.

1.2 Problem Space

1.3 Out of Scope

• Physical drone manufacturing
• Weaponization or lethal-autonomous-weapon (LAWS) capabilities — explicitly excluded
• Operations in regulated-export-controlled markets without separate ITAR/EAR review
• Fixed-wing or hybrid VTOL platforms (addressed separately if needed)

2. Decision

Adopt a hierarchical-mesh swarm topology with Raft consensus for cluster-head coordination, Gossip for environmental map dissemination, and MAPPO-based CTDE (Centralized Training, Decentralized Execution) as the MARL policy. The architecture integrates the RuView CSI sensing stack as the primary payload sensor, with OccWorld (ADR-147) as the environment prior for mission planning.

All design choices target legal civilian operations first. Dual-use swarming capability (USML Category VIII(h)(12)) requires ITAR/EAR classification review before export.

3. Swarm Architecture

3.1 Topology Selection

Hierarchical-mesh configuration:

Ground Control Station (GCS)
        │  (Sub-GHz backbone, MAVLink v2 signed)
        ▼
┌─────────────────────────────────────────┐
│          Cluster Head (CH) — elected    │
│  Role: task allocator + path planner   │
│  Runs: OccWorld prior, MAPPO centralized│
│        critic, Raft leader             │
└──────┬───────────────────────┬──────────┘
       │    (Wi-Fi 6 mesh)     │
  ┌────▼────┐             ┌────▼────┐
  │ Node A  │─────────────│ Node B  │  ... N worker nodes
  │ ESP32-S3│             │ ESP32-S3│
  │ Jetson  │             │ Jetson  │
  │ UWB     │             │ UWB     │
  └─────────┘             └─────────┘

For fleets ≥ 30 drones: form multiple clusters of 8–12 nodes; cluster heads form a peer-to-peer mesh among themselves. Each cluster operates semi-autonomously.

3.2 Consensus Protocols

Raft leader selection criteria (beyond standard Raft randomized timeout): remaining battery ≥ 60%, link quality to ≥ 2/3 followers ≥ −80 dBm RSSI, geometric centrality score (minimize max distance to any follower), onboard Jetson utilization ≤ 70%.

3.3 Communication Stack

Layer           Protocol            Band            Latency    Data Rate
─────────────────────────────────────────────────────────────────────────
Command/control  MAVLink v2 (signed) Sub-GHz 900 MHz  30–100 ms   <1 Mbps
Swarm state sync DDS (RTPS, ROS2)    Wi-Fi 6 5 GHz    <10 ms      up to 9 Gbps
CSI data (raw)   Custom UDP framing  Wi-Fi 6 5 GHz    <20 ms      ~50 Mbps/node
Relative ranging UWB (DW3000)        3.1–10 GHz       <5 ms       10 cm precision
UTM/BVLOS backhaul 4G/5G LTE         Licensed band     ~50 ms      10–100 Mbps
Long-range status LoRaWAN            868/915 MHz       ~2 s        <50 kbps

MAVLink v2 signing (HMAC-SHA256 per message) is mandatory for all inter-drone messages. TLS 1.3 for all ground-to-cloud links. DDS topics for swarm state use RTPS with RELIABLE QoS for task state, BEST_EFFORT for telemetry.

All drones use Remote ID broadcast (802.11 + Bluetooth, per FAA/EU requirements): operator position, drone position, altitude, and session ID broadcast at 1 Hz minimum.

4. Strategy Formulations

4.1 Formation Control

Three modes, selected per mission profile:

Mode F1 — Virtual Structure (precision): All nodes maintain fixed 3D offsets from a virtual reference frame propagated by the cluster head. Used for: systematic coverage grids, corridor inspection, coordinated approach. Fragile to node dropout — use when cluster is stable and mission requires geometric precision.

Mode F2 — Leader-Follower (adaptive): One drone follows a computed path; followers maintain ≥ 2 m radial offset from the leader's trajectory. Used for: linear infrastructure inspection, convoy escort. Leader failover: RAFT elects new leader from followers within 300 ms.

Mode F3 — Reynolds Flocking (emergent): Each node applies three rules with tunable weights:

• Separation: repulsion force scales as 1/d² for d < d_min (default 2.5 m)
• Alignment: weighted average heading of k = 6 nearest neighbors
• Cohesion: steering toward centroid of k neighbors

Extended with: obstacle avoidance (4th rule), OccWorld-informed zone repulsion, goal- seeking bias toward unscanned probability-map cells. Used for: large-scale area search, dynamic obstacle environments. No geometric precision guarantee.

Formation transitions (F1↔F2↔F3) are orchestrated by the cluster head based on mission phase and swarm health (dropout count, link quality distribution).

4.2 Path Planning

Primary: RRT-APF Hybrid

An RRT* planner generates globally near-optimal paths per drone. An APF (Artificial Potential Field) layer provides real-time reactive collision avoidance between planned paths. Inter-drone path intersections are treated as virtual obstacles in RRT-APF expansion (MAPF-inspired). Validated at <0.3 s computation time at high obstacle density.

Input:  OccWorld future occupancy grid (ADR-147 output)
        Current drone position (UWB + IMU fused EKF)
        Task allocation result (target cell or waypoint)

Stage 1 — RRT* global planner:
  Samples free-space voxels from OccWorld occupancy
  Builds tree; rewires for shortest path to target
  Outputs: waypoint sequence W = [w0, w1, ..., wN]

Stage 2 — APF reactive layer:
  At each timestep: compute repulsion from neighbors + obstacles
  Blend APF vector with direction to next waypoint
  Max turn rate: 30°/s; max acceleration: 0.5 m/s²

Stage 3 — Swarm clock collision check:
  Broadcast predicted path segments over DDS
  Detect spatial-temporal intersections with other drones' paths
  Insert virtual obstacle at intersection; replan affected segment

Fallback: Boustrophedon (systematic coverage) When no target is known (initial area search), each drone receives a partition of the total area from the cluster head and executes a lawnmower pattern at spacing equal to 2× CSI detection range (~28 m for the RuView Wi2SAR configuration).

4.3 Task Allocation

Auction-based with FNN scoring (hybrid):

1. Cluster head announces task T (target cell, priority, deadline)
2. Each drone computes bid b_i:
      b_i = FNN([dist_to_T, battery_pct, link_quality, csi_confidence, workload])
   FNN: 4-layer (64→32→16→8), ReLU, trained offline with Adam
   Output: affinity score ∈ [0, 1]; lower = more capable
3. Drone with lowest b_i (best fit) wins; CH broadcasts assignment
4. If winner fails to acknowledge within 500 ms, second-lowest wins

For N tasks and M drones simultaneously: solve as assignment problem. Use Hungarian algorithm for N,M ≤ 20; greedy auction rounds for larger sets.

Energy-aware constraint: drone with battery < 20% is excluded from new task bids; assigned RTH (Return to Home) or hover-as-relay role.

4.4 Coverage & Search Strategy

Phase 1 — Systematic (high-confidence sweep): Partition total area into equal-area cells across active drones. Each executes boustrophedon at flight altitude h₁ = 30 m, speed 5 m/s. CSI scan width ~28 m. Lateral overlap 20% for redundancy. No inference during transit — only at waypoints.

Phase 2 — Probability-map guided (Bayesian pursuit): Each drone maintains a shared probability grid P(x,y) of victim presence. CSI confidence scores update the grid via Bayesian rule:

P(victim @ cell) ∝ P(CSI_detect | victim present) × P(victim_prior)

Drone re-routes to the highest-entropy cell it has not yet visited. Shared grid disseminated via Gossip; cluster head resolves conflicts on write collision.

Phase 3 — Convergence (multi-drone triangulation): When P(victim) > 0.75 in any cell: cluster head assigns 3 nearest available drones to surround the cell at 3 distinct azimuth angles (120° separation). Multi-view CSI fusion via ruvector/viewpoint/attention.rs (CrossViewpointAttention) improves localization to ≤ 2 m accuracy at 3+ viewpoints.

Pheromone map (emergent coordination): Virtual pheromone field overlays the probability grid. Drones deposit pheromone on visited cells; pheromone evaporates at rate τ = 0.98/s. Pheromone steers drones away from recently scanned areas without central coordination — useful when mesh connectivity is degraded.

4.5 Emergent Behavior Policies

5. Self-Learning Integration

5.1 MARL Architecture — CTDE (MAPPO)

Training: centralized. A global critic receives full swarm state S = {positions, velocities, CSI readings, occupancy map, task queue}. N actor networks share weights (parameter sharing reduces state space curse for homogeneous swarms) and receive only local observations O_i.

Execution: decentralized. Each drone runs its actor network on local observations only; no inter-drone communication required for policy inference (communication is used for coordination, not policy inference).

Observation O_i (per drone at timestep t):
  - Own position, velocity, heading (from UWB-EKF)
  - CSI reading + confidence score (from wifi-densepose pipeline)
  - Neighbor positions within 50 m (k=6 nearest, DDS topic)
  - Probability map tile (5×5 cells centered on own position)
  - Battery level, link quality to CH
  - Current task assignment + deadline

Action A_i (continuous):
  - Δ heading ∈ [−30°, +30°] per second
  - Δ altitude ∈ [−1, +1] m per second
  - Speed setpoint ∈ [0, 8] m/s
  - CSI scan trigger (binary)

Reward R_i:
  + 10.0  for each new cell covered (first scan)
  + 50.0  for confirmed victim detection (P > 0.85)
  + 5.0   for collaborative triangulation contribution
  − 2.0   per timestep idle (encourages active coverage)
  − 100.0 for collision (d < 1.5 m to any neighbor)
  − 50.0  for geofence breach
  − 30.0  for battery depletion without RTH

Algorithm: MAPPO with shared centralized critic. Hyperparameters: lr=3×10⁻⁴, clip ε=0.2, GAE λ=0.95, entropy coefficient 0.01 (encourages exploration). Batch size 2048 transitions; 10 PPO epochs per update.

For heterogeneous fleets (e.g., CSI sensor drones + relay drones): switch to A-MAPPO (Attention-enhanced MAPPO) where attention mechanism over neighbor representations allows policy to adapt to different neighbor types.

For adversarial/anti-jamming scenarios: Use IPPO (Independent PPO) with no shared critic — fully decentralized, robust to node compromise.

5.2 Sim-to-Real Transfer

Training environment: Gazebo + PX4 SITL (Software In The Loop) with domain randomization over:

• Wind: 0–12 m/s Dryden turbulence model
• CSI noise: Gaussian noise on amplitude, von Mises noise on phase
• Motor response: ±15% thrust coefficient variation
• Communication: random 10–30% packet loss; 0–200 ms extra latency

Domain randomization distribution widths start narrow; anneal to 2× physical range over 500 training episodes to avoid reward collapse.

Sim-to-real gap mitigation: freeze MARL policy weights; use classical adaptive control (PID with integral wind-up limits) for disturbance rejection in flight. No in-flight gradient updates to the MARL policy — update only in scheduled offline retraining cycles.

5.3 SONA Trajectory Learning (In-Mission Pattern Extraction)

During operational missions, record trajectories as (O_i, A_i, R_i) triples into a replay buffer on the cluster-head Jetson (rolling 10 k transition buffer). Post-mission:

1. Filter high-reward subsequences (R > 0 for ≥ 5 consecutive steps)
2. Extract pattern fragments: (trigger_obs_embedding, action_sequence)
3. Store via mcp__claude-flow__hooks_intelligence_pattern-store
4. Retrieve similar past fragments via mcp__claude-flow__agentdb_pattern-search
   during the next mission briefing for warm-start exploration

This is the SONA analogue for drone missions: successful coordination patterns (e.g., "approach victim from 3 directions when P > 0.7") become reusable behavioral priors.

5.4 Federated Learning Across Missions

After each mission, each drone's Jetson computes a gradient update delta from its local replay buffer (no raw data leaves the drone — privacy-preserving). Cluster head aggregates via FedAvg:

θ_global ← θ_global + η × (1/N) × Σ_i Δθ_i

Updated weights broadcast to all drones before next deployment. This allows the MARL policy to improve across missions without requiring a simulation reset.

Constraint: federated update is only applied if ≥ 5 drones contributed gradients and the policy validation score (on held-out sim episodes) does not decrease by > 5%.

6. CSI Sensing Integration (RuView Payload)

6.1 Drone Payload Architecture

┌─────────────────────────────────────────────────────────────┐
│                   Drone Node (per aircraft)                  │
│                                                             │
│  ┌──────────────┐    serial/SPI    ┌─────────────────────┐  │
│  │  ESP32-S3    │ ──────────────── │  Jetson Orin Nano   │  │
│  │  8MB flash   │                  │  (40 TOPS INT8)     │  │
│  │  WiFi CSI    │  ┌─────────────┐ │                     │  │
│  │  monitor mode│  │ UWB DW3000  │ │  wifi-densepose     │  │
│  └──────────────┘  │ 10 cm range │ │  signal pipeline:   │  │
│                    └─────────────┘ │  • ADR-134 CIR/ISTA │  │
│  ┌──────────────┐                  │  • ADR-135 calibrat.│  │
│  │ Sub-GHz radio│  MAVLink v2      │  • ADR-146 RF-enc.  │  │
│  │ (command)    │◄────────────────►│  • OccWorld prior   │  │
│  └──────────────┘                  │  • MARL actor net   │  │
│                                    └─────────────────────┘  │
│  ┌──────────────┐  ┌──────────────────────────────────────┐  │
│  │  Wi-Fi 6     │  │ PX4 FMUv6X (flight controller)      │  │
│  │  (data mesh) │  │ uORB <10 ms; MAVLink; ROS2 native   │  │
│  └──────────────┘  └──────────────────────────────────────┘  │
└─────────────────────────────────────────────────────────────┘

6.2 CSI Pipeline on the Drone

The existing wifi-densepose-signal Rust pipeline runs on Jetson:

ESP32-S3 (CSI capture, 802.11n monitor mode, 56 subcarriers, 2×2 MIMO)
  ↓ serial TDM protocol (wifi-densepose-hardware)
Jetson: wifi-densepose-core → CsiFrame
  ↓ ADR-134: ISTA L1 sparse recovery → CIR (multipath profile)
  ↓ ADR-135: subtract empty-room baseline → human perturbation
  ↓ ADR-146: RF encoder multitask heads → {presence, count, keypoints, confidence}
  ↓ confidence score + 3D position estimate
Swarm DDS topic: /drone_{id}/csi/detection
  ↓
Cluster Head: Bayesian grid update + Phase 2/3 trigger

CSI scan frequency: 10 Hz during coverage, 20 Hz during Phase 3 convergence. Battery impact: ESP32-S3 in monitor mode ≈ 220 mA at 3.3 V = 0.73 W (negligible vs. ~200 W total drone consumption).

6.3 Multi-Drone Multistatic Fusion

When ≥ 3 drones are within mutual CSI link range of a target, the ruvector/viewpoint/ modules are invoked at the cluster head:

// viewpoint/attention.rs — CrossViewpointAttention
let fused = cross_viewpoint_attention(
    drone_csi_readings,  // Vec<CsiFeature> from each drone
    drone_positions,     // Vec<Position3D>
    geometric_bias,      // GeometricBias from viewpoint/geometry.rs
);
// Cramer-Rao bound: localization uncertainty ∝ 1/sqrt(N) for N independent viewpoints
// 3 drones: ~2.5× accuracy improvement vs single drone (5 m → 2 m)

The coherence_gate.rs Accept/PredictOnly/Reject/Recalibrate states gate mission decisions: a Reject state (coherence too low) prevents false positive victim reports.

6.4 OccWorld Integration (ADR-147 Output as Mission Prior)

Before deployment, the cluster head runs OccWorld inference on the last-known environmental scan of the target area:

OccWorld output: predicted 3D occupancy grid [T+1, T+5] at 0.2 m/voxel
  ↓
Extract free-space voxels → valid drone flight volumes
Extract occupied voxels (walls, debris) → no-fly zones + search targets
Assign victim-probability prior to partially-occupied voxels (rubble zones)
Feed into RRT* as obstacle map + probability-weighted goal sampling

This allows the swarm to pre-plan without real-time sensing in GPS-denied / comms- limited environments during ingress.

7. Vertical Applications

7.1 Mission Profiles (Practical → Exotic)

TIER 1 — Practical (near-term, regulatory-feasible)

P1: Search and Rescue — Structural Collapse

• Fleet: 6–12 drones; 3 CSI sensor + 3 relay/mapper
• Mission: systematic CSI sweep of rubble field; victim localization to ≤ 2 m
• Regulatory: Part 107 BVLOS waiver (US) or SORA Specific (EU); Remote ID mandatory
• Hardware: DJI Matrice 350 class body + Jetson Orin Nano + ESP32-S3 payload
• Key integration: Phase 1→2→3 coverage strategy; multistatic triangulation
• Performance target: 160,000 m² in ≤ 4 min (4-drone swarm, extrapolated from Wi2SAR)
• References: Wi2SAR (arxiv 2604.09115); wifi-densepose-mat crate (disaster MAT)

P2: Infrastructure Inspection — Power Lines / Bridges

• Fleet: 3–8 drones; formation F2 (leader-follower) along asset corridor
• Mission: simultaneous multi-angle visual + thermal + CSI anomaly detection
• Regulatory: Part 107 or BVLOS waiver per corridor; may require coordination with utility operator for airspace
• Payload: RGB + thermal camera (existing) + optional CSI for cable sag sensing
• Key integration: Mode F2 formation; 6G AI integration (arxiv 2503.00053)

P3: Precision Agriculture

• Fleet: 4–12 sprayer drones; lawnmower Phase 1 coverage
• Mission: NDVI multispectral mapping + targeted variable-rate spraying
• Regulatory: Well-established under Part 107; some states have ag-specific exemptions
• Key integration: Boustrophedon coverage; energy-aware task allocation (low-battery drones handle mapping, not heavy spraying payload)
• Note: CSI sensing not primary here; GPS precision required; RTK GPS recommended

P4: Wildfire Perimeter Monitoring

• Fleet: 6–20 drones in relay chain around fire perimeter
• Mission: continuous thermal monitoring; perimeter map update every 2 min
• Regulatory: FAA COA (Certificate of Waiver) for wildfire response; streamlined process in US; drones operate in Temporary Flight Restrictions (TFR) with waiver
• Key integration: Gossip map dissemination for shared perimeter map; LoRaWAN for long-range status back to incident command

P5: Surveying & Photogrammetry

• Fleet: 3–6 drones; virtual structure formation for overlapping coverage
• Mission: generate point cloud / orthomosaic of construction site or terrain
• Regulatory: Part 107 standard (most straightforward)
• Key integration: Mode F1 formation; boustrophedon; standard outputs to Pix4D / OpenDroneMap

TIER 2 — Specialized (mid-term, requires waivers or sector coordination)

S1: Underground Mine / Tunnel Inspection

• GPS-denied: UWB inter-drone ranging is the primary navigation reference
• SLAM: visual-inertial odometry on Jetson (VINS-Mono or Basalt)
• Fleet: 2–4 nano-class drones (sub-250g; fits tunnel diameter constraints)
• Dust/explosion rating: required for coal mines (ATEX/IECEx Zone 1 housing)
• CSI integration: CSI sensing for trapped miner detection through rock/timber
• Key constraint: comms range severely limited; Gossip over BLE mesh; no GCS link

S2: Offshore Oil & Gas Asset Inspection

• Challenge: autonomous landing on moving vessels (active compensation required)
• Fleet: 2–4 industrial-class drones with corrosion-resistant coating
• Sensor suite: electrochemical gas sensors (H₂S, CH₄); thermal; visual
• Regulatory: EASA Specific-category SORA; offshore exclusion zones; coordination with maritime traffic authority
• Key integration: Formation F2 for inspection runs; adaptive hover compensation for vessel motion (EKF with vessel IMU input via 5G link)

S3: Emergency Telecom Relay

• Fleet: 6–12 drones as flying LTE/5G repeaters after disaster
• Each drone carries a compact SDR (e.g., USRP B200mini equivalent)
• Mission: maintain coverage for first responders when ground infrastructure fails
• Flight altitude: 150–200 m AGL for maximum terrestrial coverage (~5 km radius)
• Relay chain: each drone relays to next; 6-drone chain extends coverage 30 km from GCS
• Regulatory: emergency authority coordination (FEMA/FCC in US)
• Key integration: energy-aware relay chain optimization; battery-rotation scheduling

S4: Environmental Monitoring — Air Quality / Methane

• Fleet: 4–8 drones on scheduled patrol routes; multi-day deployment with battery rotation from ground charging stations
• Sensors: electrochemical or NDIR sensors; temperature/humidity; particulate
• Data pipeline: readings aggregated to cloud time-series database; anomaly detection
• CSI integration: optional — detect worker presence in monitored zone for safety
• Regulatory: Part 107 (≤ 400 ft AGL) or BVLOS waiver for extended patrol

TIER 3 — Exotic / Advanced (long-term; active research; some regulatory hurdles)

E1: Underwater-Aerial Hybrid Swarm (Cross-Domain SAR)

• Architecture: aerial drones (above surface) relay comms for submersible drones
• Cross-domain handoff: acoustic comms underwater ↔ RF above surface
• Application: flooded structure search; open-water drowning recovery
• Key research: adaptive relay free-space networking (PMC12737092, 2025)
• Hardware gap: no production drone supports both air and water flight
• Timeline: 5–8 year horizon for operational systems

E2: Morphing / Docking Swarm Structures

• Architecture: drones physically dock mid-air to form larger rigid structures
• Application: distributed manipulation; temporary bridge segment; sensing array
• Key research: ModQuad (UPenn); 4-module airborne docking demonstrated
• Challenge: docking precision ±1 cm required; load redistribution control
• Timeline: 5+ years for ≥ 8-module practical systems

E3: Artistic / Entertainment Light Shows (Large Scale)

• Architecture: pre-programmed choreography + GPS time-sync (NOT consensus-based)
• Scale: 300–3000+ drones; growing to 10,000-unit shows by 2028 (industry projections)
• AI enhancement (current research): generative AI for choreography optimization; natural emergent motion sequences replacing rigid waypoint sequences
• Regulatory: FAA COA per show; Remote ID mandatory; pyrotechnic coordination
• Key difference from SAR: these shows use synchronized pre-programmed paths, not autonomous swarm decisions; GPS spoofing is a serious threat at this scale
• Swarm coordination applicable for: dynamic audience-responsive formation changes

E4: Bio-Hybrid Micro-Swarm (10+ year horizon)

• Concept: backpack actuators on insects (beetles, moths) + micro-drone wingmen
• Insects provide: chemical sensing beyond micro-drone capability; access to sub-cm spaces; ultra-low energy locomotion
• Micro-drones provide: guidance corrections; data exfiltration; comms relay
• Status: lab demonstrations only (UW Seattle, NTU Singapore)
• Regulatory: novel category; no existing framework
• Ethical/legal: animal welfare regulations apply to insects in some jurisdictions

E5: Swarm-Based Incremental Wireless Power Transfer

• Concept: transmitter drone array beamforms RF energy to receiver drones in flight
• Current efficiency: < 10% at 5 m (patents: USPTO 12444976)
• Practical use: extend hover endurance of stationary relay/sensor drone by 5–15%
• Full propulsion power via WPT: not viable with current physics
• Timeline: 3–5 years for incremental endurance extension; 10+ for meaningful propulsion supplement

E6: Quantum-Enhanced Swarm Optimization (Research Stage)

• Concept: quantum annealing for NP-hard task assignment at 100+ drone scale
• Current status: quantum-inspired classical algorithms (pigeon-inspired optimization, quantum-inspired APF — Nature Sci Reports 2025) outperform standard metaheuristics on formation control benchmarks
• True quantum hardware: IBM/IonQ gate-based quantum computers not yet fast enough for real-time swarm optimization; DWave annealing applicable for static assignment
• Timeline: 5–10 years before practical quantum advantage in swarm control

8. Legal & Regulatory Compliance

8.1 United States (FAA)

Swarm-as-entity gap: FAA treats each drone as an individually licensed aircraft. No waiver for a swarm as a single operational entity exists as of 2026. File per-drone COAs or waivers. Monitor BEYOND 2025 consortium rulemaking recommendations.

8.2 European Union (EASA)

No dedicated EU swarm regulation exists. Swarms fall under Specific category with SORA assessment. EASA is studying swarm-specific guidance (expected 2027).

8.3 Export Control — CRITICAL DUAL-USE FLAG

WARNING: ITAR-controlled capability. Drone swarming functions — specifically cooperative collision avoidance and coordinated multi-drone behavior — are explicitly controlled under USML Category VIII(h)(12): "Specially Designed components and parts... for unmanned aerial vehicles... [including] flight control systems with swarming capability."

Required action before commercialization or international collaboration:

1. Retain ITAR/EAR export control counsel
2. Classify each software module under ECCN or USML
3. Implement jurisdiction-based feature gating: swarming coordination features (task allocation, formation control, consensus protocols) must be gated behind export-controlled distribution controls
4. No source code repository with swarming algorithms may be public without fundamental research exemption documentation

December 22, 2025: New EAR regulations on drone equipment sourcing take effect; review supply chain for Chinese-manufactured components (COTS drone frames, FC boards).

8.4 Privacy & Data Protection

CSI sensing is an advantage: produces presence/pose without visual identification, inherently privacy-preserving for most use cases.

9. Safety Architecture

9.1 Collision Avoidance (Multi-Layer)

Layer 1 — Planning (proactive):
  RRT-APF path planning maintains ≥ 3 m inter-drone clearance in waypoints
  MAPF swarm clock: detect and resolve path intersections before flight

Layer 2 — Runtime (reactive):
  APF repulsion: activates at d < 5 m; scales as 1/d²
  Validated: 25-drone test → minimum 1.4 m maintained, zero collisions (PMC11858889)

Layer 3 — Emergency (fail-safe):
  d < 2.5 m: emergency brake + altitude separation (alternating up/down per cluster)
  d < 1.5 m: maximum divergence thrust (all motors to max away from nearest neighbor)

Layer 4 — Physical:
  Propeller guards on all drones
  Foam/compliant bumpers for close-proximity indoor operations

9.2 GPS Anti-Spoofing

Primary spoofing defense: UWB inter-drone ranging cross-check. GPS-reported position must be consistent with UWB-measured distance to ≥ 2 neighbors within ±0.5 m tolerance. Anomaly triggers: GPS data demoted to low-weight input; UWB + IMU dead reckoning promoted as primary position estimate.

Secondary: ML anomaly detection on EKF innovation sequence (XGBoost on PX4 sensor fusion path; ICCK 2025 pattern); sudden discontinuities in GPS-reported velocity or altitude flagged.

Tertiary: visual odometry (downward optical flow) as independent position reference in GPS-contested environments.

9.3 Anti-Jamming (RF)

Control link (Sub-GHz): FHSS (Frequency Hopping Spread Spectrum) on 900 MHz; 50-hop sequence; hopping rate 200 hops/s; jammer must cover full band to disrupt.

MARL anti-jamming (IPPO): each drone independently learns to adapt transmission power and frequency channel selection based on observed interference patterns (arxiv 2512.16813). Activated when RSSI drops > 15 dB below baseline.

Fallback if control link lost > 3 s: drone enters autonomous hold mode; executes last-assigned waypoints; attempts link re-acquisition for 30 s; RTH if no recovery.

9.4 Geofencing

• Geofence polygon stored onboard each drone (not fetched from GCS at runtime)
• Hard fence (immediate RTH + landing): flight authorization boundary + 20 m buffer
• Soft fence (audio/visual warning + speed reduction): flight authorization boundary
• No-fly zone database: AirMap or Airspace Link API; updated before each mission; stored locally for the mission duration (no runtime connectivity required)
• Enforcement: onboard CPU computes position relative to geofence at 10 Hz; GCS link loss does NOT disable geofencing

9.5 Fail-Safe State Machine

NOMINAL → (link loss > 3 s) → AUTONOMOUS_HOLD
AUTONOMOUS_HOLD → (link recovered) → NOMINAL
AUTONOMOUS_HOLD → (link loss > 30 s) → RTH
RTH → (battery < 15%) → EMERGENCY_LAND (nearest flat surface)
NOMINAL → (battery < 20%) → LOW_BATTERY_WARN (notify CH, no new tasks)
LOW_BATTERY_WARN → (battery < 15%) → RTH
NOMINAL → (collision imminent) → EMERGENCY_DIVERGE
EMERGENCY_DIVERGE → (safe separation restored) → NOMINAL
NOMINAL → (motor failure detected) → CONTROLLED_DESCENT

All transitions are onboard decisions; GCS acknowledgment not required for safety state changes (avoids dependency on comms link for critical safety responses).

10. Hardware Reference Stack

10.1 Baseline Bill of Materials (per drone node)

10.2 Software Stack

Flight Controller (PX4 v1.16 on Pixhawk 6X):
  - uORB topics: <10 ms internal latency
  - MAVLink v2 (signed) ↔ Jetson companion via UART/USB
  - ROS2 native via micro-XRCE-DDS
  - Custom MAVLink messages: CSI_DETECTION, SWARM_STATE, VICTIM_ESTIMATE

Companion Compute (Jetson Orin Nano, JetPack 6.x):
  - Ubuntu 22.04 + ROS2 Humble
  - wifi-densepose Rust workspace (cargo build --release)
  - MARL actor network (ONNX Runtime, INT8 quantized, <5 ms inference)
  - OccWorld Python subprocess (ADR-147; 375 ms/frame)
  - DDS swarm state bridge (FastDDS, RTPS)
  - AgentDB pattern store (local; syncs to GCS on link recovery)

CSI Node (ESP32-S3):
  - ESP-IDF v5.4 firmware
  - WiFi monitor mode; 802.11n; 56 subcarriers; 2×2 MIMO
  - TDM protocol (wifi-densepose-hardware crate)
  - Serial output at 921,600 baud to Jetson

Ground Control Station:
  - ROS2 Humble + QGroundControl
  - Swarm mission planner (custom; reads OccWorld output from ADR-147)
  - UTM client (AirMap SDK or Airspace Link API)
  - Remote ID monitor dashboard
  - AgentDB coordinator (pattern-search for mission warm-start)

11. Implementation Phases

Phase 1 — Foundation (3 months)

• Hardware integration: PX4 + Jetson Orin Nano + ESP32-S3 payload on single drone
• Validate CSI pipeline airborne: ESP32-S3 monitor mode functional at 30 m altitude
• MAVLink v2 signing: implement and test between Jetson and PX4
• UWB ranging: DWM3000EVB inter-drone ranging validated to ±10 cm at 50 m
• Geofencing: onboard enforcement; hard/soft fence working in SITL
• Remote ID: broadcast implementation per FAA/EU spec
• Single-drone MARL: train MAPPO actor in Gazebo; validate on physical drone

Exit criteria: Single drone with CSI payload operates autonomously within geofence; CSI detects human presence at 15 m range; Remote ID broadcast verified.

Phase 2 — Small Swarm (3 months)

• 4-drone swarm: PX4 SITL + Gazebo multi-vehicle; Raft consensus validated
• Formation control: F1 (virtual structure) and F3 (Reynolds flocking) implemented
• Phase 1→2→3 coverage strategy: boustrophedon + Bayesian grid + convergence
• RRT-APF path planner: integrated with OccWorld occupancy input (ADR-147)
• Auction-based task allocation: FNN scoring; assignment per §4.3
• Multi-drone CSI fusion: CrossViewpointAttention at cluster head; 3-drone triangulation
• Physical 4-drone flight test: open field; formation validation; CSI sweep

Exit criteria: 4-drone swarm covers 40,000 m² in ≤ 4 min; victim detected and localized to ≤ 5 m; zero collisions across 10 test flights.

Phase 3 — Mid-Scale Swarm (4 months)

• 12-drone hierarchical-mesh: cluster head election; Gossip map dissemination
• MARL MAPPO: centralized training complete; decentralized execution validated
• Federated learning: post-mission gradient aggregation working
• SONA trajectory pattern extraction: high-reward subsequence capture + retrieval
• BVLOS waiver application: Part 107 waiver filed (US) or SORA assessment submitted (EU)
• UTM integration: real-time position push to ADSP/USSP
• Anti-spoofing: UWB cross-check active; anomaly detection on EKF innovations
• Physical 12-drone SAR exercise: simulated rubble field; victim localization ≤ 2 m

Exit criteria: 12-drone swarm with BVLOS waiver authorization; SAR mission profile validated; ITAR/EAR classification completed by export counsel.

Phase 4 — Vertical Deployment (ongoing)

• Mission profile P1 (SAR): production-ready; first operational deployment
• Mission profile P2 (infrastructure inspection): formation F2; leader-follower
• Mission profile S1 (underground mine): GPS-denied navigation; UWB-SLAM
• A-MAPPO for heterogeneous fleets: CSI sensor + relay + mapper role types
• IPPO anti-jamming policy: deployed for contested-environment missions
• OccWorld Phase B swap: RoboOccWorld integration when code releases (~Q3 2025)

12. Consequences

12.1 Positive

• Directly extends the RuView CSI sensing stack to airborne deployment, unlocking the MAT (Mass Casualty Assessment Tool) crate's disaster-response mission
• Hierarchical-mesh with Raft provides production-grade fault tolerance without the O(n²) overhead of BFT
• CTDE MARL allows optimal cooperative behavior during training while keeping each drone's runtime fully autonomous (no inter-drone comms required for policy inference)
• SONA pattern extraction creates a self-improving mission library across deployments
• OccWorld occupancy prior (ADR-147) gives the path planner a physics-grounded environment model; reduces exploration time in complex environments

12.2 Risks & Mitigations

12.3 Open Issues (Forward to ADR-171)

• Cosmos WFM offline training data generation (deferred from ADR-147) — ADR-171
• Fixed-wing hybrid platform support (endurance missions) — future ADR
• Underwater-aerial cross-domain handoff protocol — future ADR
• Quantum-enhanced task assignment (E6) — future ADR when hardware matures

13. Research Notes & References

Primary Papers

Regulatory References

Evidence Quality Assessment

14. Implementation Progress (2026-05-30)

Crate wifi-densepose-swarm implemented at /home/ruvultra/projects/RuView/v2/crates/wifi-densepose-swarm/.

Milestone Status

Overall: ~98% — only M3's hardware-gated 15% (physical ESP32-S3 CSI capture) remains.

M4 — Real GPU Training (added 2026-05-30)

The MARL trainer now does genuine gradient descent via Candle 0.9 autodiff (marl/candle_ppo.rs, feature train, optional cuda):

• CandleActorCritic (64→128→64 MLP), CandleTrainer with GAE + clipped surrogate + real optimizer.backward_step(). CPU or CUDA (local RTX 5080 / GCP L4).
• A-MAPPO heterogeneous-role attention (marl/role_attention.rs): relay attention floor, role-segmented pools, sensor-gated triangulation-geometry penalty, role embeddings.
• train_marl binary: cargo run --features train,cuda --bin train_marl.
• Right-sized launch: scripts/gcp/provision_marl.sh (L4 / g2-standard-16, ~$1.40/hr — MARL is rollout-bound, not matmul-bound; A100×8 reserved for OccWorld world-model training) + run_marl_train_local.sh (local 5080).
• Verified: 5-episode CPU run shows value_loss decreasing (critic learning) + safetensors checkpointing.

Verified Benchmark Results (criterion, release mode)

Test Coverage

• --no-default-features: 67/67 tests pass
• --features itar-unrestricted: 79/79 tests pass
• Criterion benchmark harness: 4 benchmarks active

ITAR Compliance

All swarming coordination features (formation, Raft, task allocation) are gated behind #[cfg(feature = "itar-unrestricted")] per USML Category VIII(h)(12). Default builds compile and export clean stubs returning Err(SwarmError::Security(...)).

GitHub Issue

Implementation tracked at: https://github.com/ruvnet/RuView/issues/861

ADR authored with research support from ruflo-goals:deep-researcher (2026-05-30). Implementation progress tracked by ruflo-goals:horizon-tracker. OccWorld integration basis: ADR-147. Next: ADR-171 (Cosmos WFM offline data generation; renumbered from ADR-149).