Most "AI drones" send data to cloud APIs for decisions. 2-10 second latency. Requires connectivity. Single point of failure. Per-token cost at scale.
Language models hallucinate. They're non-deterministic — same input, different output. A drone that "sometimes" avoids obstacles is a drone that crashes.
Running inference on-device needs expensive GPUs. $10K+ per drone. 10-30W power draw. Heavy. Impractical for small autonomous platforms.
The alternative: massive if/else trees baked into firmware. Can't change missions without reflashing. No adaptability. Brittle at edge cases.
Sensor Data → LLM reasons → Hope for good output
(slow, expensive, hallucinates, needs cloud)
Sensor Data → 5 layers of CODE → Direct action
(fast, free, deterministic, on-device)
Each layer performs one specific reasoning task using pure Python + YAML configuration. No neural networks. No probabilistic inference. Just structured, auditable code.
For drones, L1-L5 handle all reasoning. L6 (text generation) and L7 (privacy/GDPR) can be skipped entirely. Internal benchmarks: <3ms full pipeline on M4. Targeting <100ms on Raspberry Pi 5.
Battery, GPS, obstacle distance, wind, IMU — converted to text query that L1 can pattern-match. "battery critical 8% | gps fix 12 satellites | obstacle detected 1.5m"
L1 classifies severity. L2 scores battery level, obstacle proximity, GPS status as dimensions. L4 walks a YAML decision tree: battery > 0.8 → RETURN_TO_BASE. All in code. Zero ambiguity.
Decision maps directly to motor commands. RETURN_TO_BASE → throttle 80%, heading home. HOLD_POSITION → hover. No text parsing. No interpretation. Direct execution.
Drone flies waypoints A→B→C→D in a loop. Battery drains. GPS jitters. Wind varies. MDIL handles everything.
Kill GPS. Drain battery. Spawn obstacles. Strong wind. Press them mid-flight and watch MDIL's decision tree react instantly.
Trigger a high-priority target intercept. The drone pushes through disruptions to reach the target. Battery doesn't matter. Mission first.
All 7 layers operational. Configuration-driven via YAML — any domain, any use case. Symbol Security patent filed. The architecture is stable and extensible.
L1-L5 reasoning. Sub-100ms decisions. YAML mission configs. External gates for human-in-the-loop. Targeting Raspberry Pi 5 deployment.
MDIL's Project system enables multi-instance orchestration. External Gates can coordinate between drones. Fleet-wide mission changes via YAML.