EDGE AI · GROUND SERVERS · ATHENA AI

Intelligence in the air. Analysis on the ground.

AEOS runs two layers of intelligence. In the air, the Jetson Orin NX processes every camera frame in real time — flagging disease, weeds, and obstacles before the data ever reaches the ground. After landing, your ground servers take over: orthomosaics are stitched, NDVI maps are generated, and the AI server runs deep crop analysis. One flight. Two layers. Full picture.

Request Intelligence Demo Sensor Stack
AEOS INFERENCE PIPELINE
> ONBOARD · Jetson Orin NX · 100 TOPS oak_d_pro: depthai 3.1.0 · Myriad X active yolo: 3 Hz · YOLOv6-nano · 80 class voxel_map: 3 Hz · inner 0.25 m / outer 0.5 m path_plan: A-Star · 26-connected · ~100 ms athena: keyword <50 ms | local_llm | claude telemetry: 5–10 Hz MQTT · HMAC-signed > GROUND · Dell R640 · AI Server odm: orthomosaic · NDVI maps · 3D mesh ppk: post-processing · cm geoaccuracy ai_server: deep crop analysis · model training

See your crops the way the drone sees them.

Live video from the OAK-D Pro camera streams to the ground station during every flight. What you see in the dashboard is exactly what the drone sees — AI detection overlays included, in real time.

AEOS live camera — surveying crop rows from flight altitude
AEOS Front Cam · Live JPEG snapshot stream · OAK-D Pro · mission in progress over agricultural field

Disease, weeds, pests — flagged in flight, not in post.

The onboard neural network identifies crop problems frame by frame at altitude. No cloud, no internet, no waiting. Detections appear in the ground station seconds after the drone spots them — with GPS coordinates attached.

CROP SCOUTING

Disease & Weed Signatures

80-class YOLOv6-nano model identifies disease patterns, weed presence, and vegetation anomalies at 3 Hz. Each detection is tagged with GPS-referenced lat/lon and 3D spatial coordinates.

SEMANTIC SAFETY

Human & Animal Priority

Person detections receive maximum semantic cost (10) and trigger a stop override regardless of depth readings — YOLO beats geometric sensors when class cost ≥ 7 or detection is closer than the rangefinder.

DETECTION FEED

Live Thumbnail Stream

Ground station receives detection events with JPEG thumbnail crops, confidence score, class label, and 3D position. Up to 100 detections are held in the DetectionFeed panel for operator review.

YOLO SEMANTIC COST TABLE
ClassSemantic CostInflate RadiusBehaviour
person10 (max)3.0 mSTOP override — beats all depth readings
car / truck / bus8–94.0–5.0 mSTOP if within stop_distance_m
dog82.0 mHigh-priority avoid
cat / bird5–71.5 mLateral shift or slow
static structures1–31.5 mGeometric avoidance via voxel map
default (unknown)41.5 mStandard margin
YOLOv6 apple detection — bounding boxes with class labels and confidence scores on apple tree
Field result · YOLOv6 · Apple detection · Myriad X onboard inference
YOLOv6 pest damage detection — bichado (codling moth damage) identified with bounding boxes
Field result · YOLOv6 · Pest damage (bichado) · confidence scored detections

It builds a 3D map of your field as it flies.

Depth sensors and camera frames are fused in real time into a live 3D occupancy map the drone uses for safe navigation. Obstacles are identified and avoided before the operator even sees them on the dashboard.

TWO-ZONE VOXEL GRID
ZoneVolumeResolutionMemoryUse
Inner (drone-centred) 20 × 20 × 10 m 0.25 m 0.49 MB Reactive avoidance, person detection, doorway resolution
Outer (world-fixed) 100 × 100 × 50 m 0.5 m 7.63 MB A-Star path planning, mission routing
UNKNOWN = 0
FREE = 1
OCCUPIED = 2
Ray-cast to FREE
Auto-save every 60 s
Map reuse within 50 m of home
A-STAR PATH PLANNER
  • 26-connected 3D neighbours (face + edge + corner)
  • Euclidean heuristic · searches FREE voxels only
  • OCCUPIED and UNKNOWN cells blocked (configurable)
  • Obstacles inflated by 1.5 m safety margin before planning
  • Semantic cost overlay: g(n) = distance + semantic_cost
  • Corridor detection via distance-transform medial axis
  • ~100 ms on 200 × 200 × 100 outer grid with 10 % occupied
> /nav/mode: SURVEY | INSPECT | CORRIDOR | RTL SURVEY: high altitude · terrain follow · wide margins INSPECT: low altitude · 360° scan per waypoint CORRIDOR: A-Star routes · lateral avoidance through rows RTL: max safety · conservative thresholds

Look before you fly into tight spaces.

Before entering crop rows or confined areas, AgroExplorer hovers and scans all directions, building a complete picture of what surrounds it. It only proceeds when the path is clear.

SCAN360 SEQUENCE
8 headings

45° increments · 27° overlap between captures (72° HFOV vs 45° step) = full 360° coverage guaranteed.

~24 s total

Each step: rotate + settle + depth-capture → 3 s per heading. Drives yaw via MAV_CMD_CONDITION_YAW (param 115).

Voxel integration

Every capture integrates the current OAK-D depth frame into the VoxelMap. State machine: idle → rotating → settling → capturing → complete.

INSPECT missions

Optional 360° scan fires before each inspection waypoint to ensure full local coverage before close-in approach.

SENSOR FUSION SOURCES

OAK-D Depth

10 Hz · 3D voxel primary

OAK-D YOLO

3 Hz · semantic costmap

TFMini Forward

10 Hz · FORWARD sector

TFMini Ground

10 Hz · terrain AGL

HereFlow

10 Hz · EKF flow

UM982 RTK

5 Hz · world anchor

Pixhawk EKF

20 Hz · NED pose

9-Sector Map

25 Hz · reactive

From raw flight data to NDVI maps, orthomosaics, and 3D canopy models.

Every mission generates survey data that the ground processing server turns into actionable outputs — georeferenced maps, vegetation health indices, and 3D crop structure — ready after landing.

ORTHOMOSAIC

RGB Field Map

The ELP 16MP downward camera captures high-resolution RGB frames across the entire survey area. OpenDroneMap stitches them into a georeferenced orthomosaic at 3–7 mm/px ground sampling distance — every row visible, every anomaly located.

NDVI · GNDVI · SAVI

Vegetation Health Indices

A NIR-modified second camera (850nm longpass) pairs with the RGB camera to compute NDVI and GNDVI across the whole field. Healthy vegetation maps green; stressed or bare zones show yellow and orange — guiding targeted intervention.

3D CANOPY

Stereo Depth Structure

OAK-D Pro stereo depth maps canopy height, volume, and row gaps in 3D during flight. Plant height anomalies (lodging, stunted growth) are flagged automatically. Canopy volume × NDVI gives an onboard biomass estimate.

AGRICULTURAL INDEX REFERENCE
IndexFormulaWhat It DetectsCameras Needed
NDVI(NIR − Red) / (NIR + Red)Overall vegetation healthRGB + NIR
GNDVI(NIR − Green) / (NIR + Green)Chlorophyll in dense canopyRGB + NIR
SAVI1.5 × (NIR − Red) / (NIR + Red + 0.5)Vegetation health on bare soilRGB + NIR
ExG2×Green − Red − BlueExcess green proxy (RGB only)RGB
NDRE (Phase 4)(NIR − Red Edge) / (NIR + Red Edge)Early nutrient stress, precise chlorophyllRGB + NIR + Red Edge

A 3D record of every mission.

After landing, AEOS stitches keyframe imagery into a coloured 3D mesh — a permanent visual record of the field you can navigate, annotate, and share with your agronomist. Coming in Phase 6 of the AEOS roadmap.

KEYFRAME CAPTURE

Triggered on motion

RGB frame + camera pose captured when drone moves >1 m OR rotates >10° OR >50 new voxels occupied OR every 5 s. JPEG Q80 at 416 × 416 ≈ 30 KB/frame. ~200–400 keyframes per 10-minute flight.

COLOURED VOXELS

Parallel RGB grids

Inner 80 × 80 × 40 × 3 = 0.73 MB, outer 200 × 200 × 100 × 3 = 11.4 MB. Delta stream extended to 12 bytes/voxel (x, y, z, state, r, g, b). Coloured voxels visible live in the 3D map viewer.

POST-FLIGHT MESH

Marching cubes PLY/OBJ

MeshBuilder loads voxel grid + keyframe manifest, Gaussian smooth (σ=0.5), marching cubes → triangles, project vertices to best keyframe, sample RGB. ~20–50 K triangles; ~5–15 s compute. Three.js viewer in ground station.

Draw your survey on the map. The drone does the rest.

Click waypoints on a live satellite view, set altitude and speed, and submit. AgroExplorer validates the plan onboard and executes it to the metre — covering every row, every time.

AEOS Mission Planning — vineyard survey with waypoints placed on live satellite map
Mission Planning page · AgroExplorer One connected · Waypoints placed over vineyard field
NAVIGATION MODES

Three autonomous modes, one platform

SURVEY

High-altitude orthomosaic and NDVI/NDRE grid coverage

Wide safety margins · terrain following · automated lawnmower path
INSPECT

Low-altitude close-inspection with 360° scan at each waypoint

Tight margins · OAK-D depth on full alert · hover + multi-spectral capture
CORRIDOR

A-Star planned routes through crop rows, alleys, and confined passages

Lateral avoidance · path-planned intermediate waypoints from voxel map

An AI co-pilot that knows your field.

Ask Athena what the drone is doing, what it detected, or what to do next. It answers from live telemetry instantly — and escalates to a full AI model for complex questions without any delay to the operator.

THREE-TIER REASONING
Tier 1 · <50 ms

Keyword matching for operator commands and status queries

e.g. "battery?", "RTL now", "what mode?" — instant response, no model call
Tier 2 · Local LLM

On-device language model for complex agronomic questions and mission planning

Runs on Jetson Orin NX · no external API required · low-latency reasoning
Tier 3 · Claude API

Claude cloud API for deep agronomic analysis, anomaly explanation, and full-context answers

Operator-configurable · authority levels: waypoints / commands / missions / safety
ATHENA CAPABILITIES
  • Voice input (push-to-talk) and text chat in the operator console
  • Real-time telemetry awareness — knows battery, GPS, link state, mission progress
  • Reads and generates mission YAML for waypoint planning requests
  • Action approval flow — proposes commands for operator confirmation before execution
  • Health alerts: audio beep at battery 15 % warn, 5 % critical
  • EMCON integration — Athena respects emission-control restrictions
  • Configurable authority levels: operator decides what Athena can autonomously action
  • Separate WebSocket to Athena service on port 8200