For Government & Defense

Vendor-agnostic drone swarm training. Mission-ready.

The same training pipeline that powers commercial fleets runs on air-gapped GPU infrastructure for federal programs and prime contractors — with measured sim-to-real transfer, full vendor independence, and one-click export to PX4, ArduPilot, ONNX, and ROS 2.

Mission Profiles
  • Persistent ISR — multi-drone reconnaissance at scale
  • Perimeter defense — learned threat detection & intercept
  • Search & rescue — coordinated area coverage
  • Contested logistics — resupply through GPS-denied / EW environments
  • Counter-UAS swarm experimentation
  • Operator training — HITL drills against scripted adversary swarms
Defense-Grade Capability
  • Air-gapped GPU deploy — cloud, on-prem, or classified networks
  • One-click export to PX4 (MAVLink), ArduPilot (pymavlink), ONNX, ROS 2 (lifecycle node + action server)
  • Sim-to-real loop with measured RMSE, per-axis drift rate, and Jensen-Shannon divergence
  • Differentiable physics for system identification from real PX4 / ArduPilot flight logs
  • Open architecture — REST + WebSocket APIs, JSON config, MOSA-aligned
  • Signed model exports — Ed25519 default, AWS KMS for FedRAMP / CUI
Closed-loop sim‑to‑real

Measure the gap. Tune. Re-fly. Re-measure.

Upload a real PX4 .ulg or ArduPilot .tlog. NeuHive replays the commanded inputs through the simulator and quantifies the divergence. Then it tells you what to change.

RMSE
Trajectory root-mean-squared error in meters, plus per-axis breakdown.
Drift Rate
Per-axis drift in m/s — quantifies bias rather than noise.
Jensen-Shannon
Divergence on position and velocity distributions.
Each metric drives ranked domain-randomization recommendations. One-click apply pushes the change back into the environment builder. Then retrain, re-fly, re-measure.
Environment builder

Thirteen panels to shape your training environment.

Configure mission-specific terrain, weather, threats, and sensors without writing code. Every change renders into a live 3D preview running the same engine the policy actually trains in.

Physics
Gravity, drag, motor dynamics, ground effect, rotor wash, AABB collision
Weather
Dryden turbulence + discrete gust events
Sensors
Per-channel noise, bias, dropout; GPS modes incl. degraded / spoofed / denied
Comms
Packet loss, latency distribution, neighbor topology cap
Electronic Warfare
Named threat zones with jamming radius + GPS-denied regions
Domain Randomization
Seven axes (mass, drag, motor BW, sensor noise, latency, wind, spawn jitter)
Patrol Zones
Mission-area coverage objectives
Rewards
Eight composable reward components, per-skill weights
Sim2Real
Upload .ulg / .csv flight log, replay through simulator, measured gap
Difficulty
Five-stage progression with auto-promotion gates
Training
Twenty PPO hyperparameters in the UI, no code required
Assets
Box, Wall, Cylindrical Tower, Tunnel, Bridge, Pipe, Container, Scaffolding + custom GLB / USD / FBX uploads
Scene
Spawn, Goal, Arena, Patrol Zones, EW Zones in a live 3D preview using the same physics as training
Architecture & deploy

Open. Portable. Air-gapped if you need it.

NeuHive runs on any CUDA-capable GPU infrastructure — cloud, on-premise, or on a fully air-gapped classified network. No external dependencies at runtime. No phone-home telemetry. Standard interfaces all the way down so the platform never becomes the lock-in.

  • REST + WebSocket APIs for everything
  • JSON configuration end-to-end
  • Signed model exports (Ed25519 / AWS KMS)
  • Two policy backbones: flat MLP and Queen hierarchical
  • Reference deploy: single g5.xlarge for development; horizontal scaling for production training
Export targets
ONNX Universal
INT8 / FP16, graph optimization, dynamic batch
PX4 Autopilot
MAVLink integration, SITL, Offboard mode, failsafe wired
ArduPilot
Companion package via pymavlink runner, GUIDED mode, SITL bridge
ROS 2 Package
Lifecycle node, action server, Humble + Iron, launch files
Train once. Deploy to whatever flight controller the hardware ships with.

Ready to evaluate?

Send us one PX4 .ulg or ArduPilot .tlog from a real test flight. We'll run it through the sim-to-real pipeline and walk you through the measured gap on a follow-up call.