Bee Edge AI

Program the Bee camera to collect data from the physical world you want

Build and deploy custom AI workloads on the Bee's edge computing platform. Write Python modules, push them OTA (over the air) to Bee devices, and stream results to your cloud.

Overview

Edge Modules are software programs that run on the Bee camera device to collect the data you want from the physical world. They have access to all onboard sensors and can run custom ML models alongside the native Map AI stack.

Capabilities

Feature	Details
Sensors	12.3MP camera, stereo depth imagery, GPS, IMU, accelerometer
Compute	5.1 TOPS NPU, runs inference offline
Deployment	OTA via Bee Maps infrastructure
Targeting	Country, state, metro
Output	Structured JSON e.g. detected objects Imagery Imagery and Depth Video Telemetry

Sensor Access

Edge Modules can access all onboard sensors:

Sensor	Data
Camera	12.3MP RGB frames at 30 FPS
Depth	Stereo depth imagery with distance estimation
GPS	Latitude, longitude, altitude, speed, heading
IMU	Accelerometer and gyroscope (6-axis)

Running Custom Models

There are two approaches to running custom models.

Model Type	Description
Detection Models	Models that detect and position objects in a scene: "speed limit sign exists at these coordinates"
Classification Models	Models that perform classification tasks: "is there a baby stroller in the image?"

Geographic Targeting

Deploy modules to specific regions using the Bee Maps console.

Targeting Options

Level	Description
Country	All devices in a country
State/Province	All devices in a state or province
Metro	All devices in a metropolitan area
City	All devices in a given city e.g. Santa Monica, CA

Devices receive your module only when operating within targeted regions.

Data Offload

All Edge Module output streams to your cloud via Bee Connectivity Services. Alternatively, you can upload to Bee Maps and then use our API to consume the data.

Connectivity Channels

The Bee uses two offload channels to optimize bandwidth and latency:

Channel	Use Case	Behavior
LTE	Real-time critical data, small payloads	Always on, immediate delivery
WiFi	Bulk imagery, large payloads	Batched delivery when connected (typically overnight)

Output Configuration

Configure what data gets sent and when.

Data Type	Typical Size	Default Delivery
Detections (JSON)	~1 KB per event	Real-time via LTE
Frame crop	~50 KB	Real-time via LTE
Full frame (12.3MP)	~2 MB	Usually Batched via WiFi
Depth crop	Variable	Batched via WiFi
Video clip	Variable	Batched via WiFi

Deployment Workflow

1. Create Module — Define your module configuration and upload your model via the Bee Maps console

2. Configure Output — Set your endpoint and select which data types to include

3. Set Targeting — Define geographic regions using the targeting UI

4. Staging Deploy — Push to a small device subset to validate accuracy and recall

5. Production Deploy — Roll out to your full target region via OTA

Example Use Cases

Retail & Places Churn

Monitor storefronts to detect business changes—new openings, closures, rebrands.

What you detect:

• "For lease" signs

• Changed storefront signage

• Boarded windows

• New business openings

Output: Structured change events with imagery, fed into places databases to keep POI data fresh without manual surveys.

---

Complex Intersection Video

Capture video clips at specific intersections for traffic analysis, urban planning, or safety studies.

How it works:

• Define target intersections via GeoJSON or automatically detect complex intersections based on detection count of traffic lights

• Trigger recording when devices enter the zone

• Collect multi-angle footage as different vehicles traverse the same intersection over time

Output: Geotagged video clips from multiple perspectives, timestamped for temporal analysis.

---

Long-Tail Event Capture

Detect rare but critical events that matter for autonomous vehicle training and safety validation.

Example events:

• Pedestrians with strollers in roadway

• Wheelchair users crossing

• Animals in road

• Unusual vehicle types (oversized loads, emergency vehicles)

• Construction zone edge cases

• Adverse weather conditions

How it works: Run lightweight classifiers on-device. Upload only when target events are detected. Build datasets of real-world edge cases at global scale.

Output: Annotated imagery and video of rare events, with full sensor context.

---

World Model Training Data

Collect synchronized video + depth + IMU data for training world models and vision foundation models.

Data captured:

• High-resolution video (12.3MP @ 30fps)

• Imagery and stereo depth imagery pair

• Full IMU telemetry (accelerometer + gyroscope)

• Precise GPS positioning

Targeting options:

• Specific road types (highways, urban, rural)

• Weather conditions

• Geographic regions

• Time of day

Output: The raw ingredients for building physical world simulations—synchronized multimodal sensor data at scale.

---

Getting Started

Ready to deploy custom AI workloads on the Bee platform?

• Email Us — Drop us a note to create your first Edge Module