An AI co-pilot that monitors, reasons, and acts — so you can focus on the field.
Athena is the AI assistant built into AEOS. It listens to live drone telemetry, answers your questions in natural language, warns you before problems become incidents, and can execute mission commands — with your approval. Ask by text or voice. Get answers in under two seconds.
What Athena can do for you.
From instant status reads to building a new mission mid-flight — Athena bridges natural language and drone control.
Answer Any Question
"What's the battery?" "How many waypoints left?" "What did it detect over the north corner?" Athena reads live telemetry and the mission knowledge store to answer in plain language — instantly.
Watch Proactively
Athena monitors battery drain rate, GPS quality, altitude deviation, and detection clusters in the background — and alerts you before they become problems. You don't have to ask; it tells you.
Execute Commands
Pause the mission, insert a waypoint, trigger RTL, or build a new survey plan — all via natural language. Actions that affect the drone require your explicit approval before executing.
Push-to-Talk Interface
Hold the mic button, speak, release. Whisper STT transcribes your words, Athena reasons, Piper TTS speaks the response. Full voice round-trip in ~2–3 seconds, with no cloud dependency for flight-critical paths.
Query Mission History
"What did we find last time we surveyed this field?" Athena queries the PostGIS knowledge store for prior detections, maps, and flight summaries — giving each new mission the benefit of every previous one.
Build Missions From Description
"Plan a survey of the north field at 25 metres, 80% overlap." Athena generates a YAML mission definition and uploads it to the drone for your approval — no manual waypoint placement required.
Three tiers of reasoning. The right one fires automatically.
Athena routes every query to the fastest tier that can handle it — so simple commands are instant and complex questions get full AI reasoning.
Tier 1 — Instant (< 50 ms)
Keyword pattern matching. No LLM invoked. "Battery", "status", "altitude", "pause", "resume", "RTL" are handled entirely in memory — critical for safety-time commands where a two-second delay is unacceptable.
Tier 2 — Local LLM (~1–2 s)
Qwen2.5-7B-Instruct Q8 running on the AI Server via llama.cpp. Full tool-use reasoning. Always available offline — no internet required. Handles complex queries, mission planning, and anomaly analysis when cloud is unavailable.
Tier 3 — Claude API (~3–5 s)
Best-quality reasoning for complex natural language requests. Used when internet is available. Falls back to Tier 2 automatically if connectivity drops mid-mission — zero interruption to the operator.
It watches the drone so you don't have to.
Athena's event monitor runs continuously in the background — tracking telemetry buffers, running anomaly checks every second, and alerting you the moment something needs attention.
| Trigger | Condition | Response |
|---|---|---|
| Battery drain anomaly | Drain rate > 1.5× rolling average | Alert + estimate time remaining |
| Battery critical | < 25% while executing | Alert + recommend RTL |
| GPS degradation | Fix type drops below 3D | Alert |
| GPS lost | No fix for > 5 s | Alert + recommend pause |
| Altitude deviation | > 5 m from planned | Alert |
| Progress stall | No waypoint reached in 60 s | Alert + diagnose |
| Detection cluster | 3+ detections within 50 m | Alert + recommend investigation |
| High-confidence detection | Single detection > 90% | Alert with location + class |
| Failsafe triggered | Any failsafe flag goes true | Immediate alert |
You stay in control. Athena asks before it acts.
Every action Athena can take has an authority level you configure. By default, anything that affects the drone requires your explicit approval — a single click in the chat panel before the command fires.
Always Allowed — No Approval
Read telemetry, query mission state, check safety status, review detection history, look up prior field data. Athena reads freely and answers questions. Nothing changes on the drone.
Always Allowed — Push Alerts
Send alerts, status summaries, and mission recommendations to the chat panel. Athena tells you what it sees. You decide whether to act.
Requires Approval by Default
Insert waypoints, send drone commands, start missions. Athena proposes the action with full context in the chat panel — you click Approve or Deny. 30-second timeout cancels the action if you don't respond.
Auto-Execute by Default
Emergency RTL on critical battery, pause on geofence approach. Safety overrides execute automatically — configured to auto by default because waiting for approval is unsafe. Configurable per deployment.
| Action Category | Options | Default |
|---|---|---|
| Waypoint creation / insertion | Auto · Approval required · Disabled | Approval required |
| Drone commands (pause / resume / RTL) | Auto · Approval required · Disabled | Approval required |
| Mission start / upload | Auto · Approval required · Disabled | Approval required |
| Safety override (emergency RTL / pause) | Auto · Approval required | Auto |
Speak to your drone. Hear it speak back.
Push-to-talk voice interface with on-premises speech processing — no cloud dependency for transcription or synthesis during flight operations.
| Component | Technology | Latency |
|---|---|---|
| Voice Activity Detection | silero-vad (CPU) | ~5 ms |
| Speech-to-Text | Whisper large-v3 (GPU) | ~0.5–1 s per 5 s audio |
| Reasoning | Tier 1 keyword or Tier 2 LLM | <50 ms – 2 s |
| Text-to-Speech | Piper TTS — en_US-amy-medium (CPU) | ~200 ms |
| End-to-end round trip | ~2–3 s typical |
- Hold the mic button in the Athena panel → browser captures audio
- Audio streams to the AI Server via WebSocket as 16 kHz PCM
- silero-vad detects the speech segment boundary
- Whisper large-v3 transcribes the segment on GPU in ~1 second
- Transcribed text enters the Intent Router — Tier 1 or Tier 2 fires
- Piper TTS converts the response text to speech on CPU in ~200 ms
- WAV audio returns to the browser and plays automatically
- Full round trip runs on the AI Server — no external voice APIs required
Where Athena runs.
Athena lives on the AI Server — the same machine that handles ML training, crop analysis inference, and model management. It connects to the drone via MQTT and to the operator via the ground station browser.
Being honest about the boundaries.
- No image analysis during flight — YOLO detections come from the Myriad X on the OAK-D Pro. Athena sees detection results, not camera frames.
- No direct MAVLink access — all commands route through the Orchestrator API → MQTT → drone. Athena never talks to the flight controller directly.
- No map tile serving — post-flight orthomosaics and NDVI maps are generated by the Dell R640 processing server. Athena can query their metadata and results.
- No vision model — Athena is a text-reasoning AI. It works from telemetry data and detection events, not raw camera imagery.
See Athena running a real mission.
During any AgroExplorer demonstration, Athena is live in the ground station — monitoring telemetry, answering questions, and managing alerts in real time. Ask it anything during the flight.