Machines that act on what they see.
We build intelligent physical systems: computer vision, precision mechatronics and safety-critical embedded control, engineered to work at distance, in weather, unattended. Three divisions, one platform underneath all of them.
Wildlife Systems — Goose Deterrent System, Model 1 Rev C. One of three divisions.
One platform, three problems.
Each division owns its own market, hardware and compliance surface. All three run on the same perception, motion and safety stack.
Wildlife Systems
Active — autonomous detection and non-contact deterrence for outdoor property. First product: the Goose Deterrent System, Model 1, in field test. See the division
Site Monitoring
In development — autonomous observation of large outdoor properties, where the value is in noticing rather than acting.
Inspection & Sensing
In development — precision-pointed sensor payloads on fixed outdoor mounts, reusing the same calibration and motion work.
Divisions marked in development are engineering directions. They are not products and are not available.
Watch it decide.
The model tracks geese and looks for people in the same frame. A person anywhere in view cuts the emitter before an aim command is ever issued.
Illustration of the detect–track–deter sequence and the human-detect interlock. Not a live feed.
Three million geese that never leave.
Canada geese used to migrate through. Now roughly 3.2 million live in the United States year-round, and they settle on exactly the places people maintain: golf courses, retention ponds, campus lawns, farm fields.
A single adult produces two to four pounds of droppings a day. On turf that means cleanup labor without end. In water it means phosphorus, and phosphorus means algae, and algae collapsing means the dissolved oxygen fish need is gone.
The usual answer is a person walking out to chase them off. It holds for an afternoon. Decoys, reflective tape and noise cannons hold until the flock learns they're harmless. None of it produces a record of anything.
This program started on a private lake that was losing its dissolved oxygen, with a flock that came back every morning.
The Goose Deterrent System
It watches the property, identifies goose activity, traces a 532 nm light pattern around the birds inside an approved zone — no physical contact — and logs every event so you can prove whether it worked.
Detection cycle
Detection mAP@50
Emitter wavelength
Measured on the current model, Revision C, under controlled conditions.
Three hard problems, solved once.
Every division inherits the same stack. That is the point of the company — the second and third products do not start from zero.
Perception at distance
Detection running locally on embedded hardware, tuned for small targets against water glare, moving vegetation, low light and fog.
Precision actuation
Closed-loop pan and tilt with low backlash and repeatable homing, so an image coordinate becomes the same physical aim point every time.
Safety-critical control
A dedicated microcontroller owns the limits, interlocks and the emitter enable line. The perception computer can request. It cannot authorize.
Wildlife Systems
Autonomous detection and non-contact deterrence for outdoor property. The division's first product is the Goose Deterrent System — Model 1 — currently in field test.
How it works
Detect
A dome-protected camera on a full-azimuth rotating drive scans the designated area.
Classify
The onboard model separates geese from terrain, water, vegetation and equipment — and from people. Every frame is checked for humans, not just for birds.
Track
The system selects a target and holds a position estimate as the bird moves.
Aim
Image coordinates become pan and tilt angles through the calibrated camera-to-emitter geometry, then get checked against the site's approved zone.
Deter
The safety controller verifies zone, state, confidence, hardware health, comms and a clear human channel before enabling a duration-limited pattern around the target.
Record
Detections, deterrence events and response latency are logged for effectiveness analysis.
Inside the housing
Model 1, Revision C — the build currently in field test.
- A
Camera
High-resolution optics in a protective dome, aligned to the emitter axis.
- B
Aim motors
Closed-loop stepper drives for pan and tilt of the optical head, with endstops and homing.
- C
Rotating motor
Azimuth drive carrying the whole assembly through its scan pattern.
- D
Control board
Edge compute plus the independent safety microcontroller that owns the emitter enable line.
- E
Emitter
Heat-sinked 532 nm optical head, hardware-interlocked and duration-limited in firmware.
Where it goes
Golf courses
Ponds and adjacent turf, covered before play. Operating zone excludes tees, paths and clubhouse sightlines.
Farms
Crop blocks, ponds and irrigation areas during the hours birds arrive. Usually off-grid, which drives the solar configuration.
Retention ponds & private lakes
Shoreline congregation, with before-and-after activity data lake managers can use.
Corporate campuses
Walkways, entrances and lawns. Foot traffic means tighter zones and heavy reliance on the human-detect interlock.
Industrial & construction sites
Controlled work zones that already have restricted access and defined boundaries — the simplest safety case.
Municipal & aviation
Long-term objectives. Open public access and airside operations raise the bar on regulation, certification and liability, and we don't represent the system as approved for either.
Revisions
Model 1 is on its third built revision. Each one was driven by what broke or fell short in the last.
- Built
Rev A · Bench proof
Vision, motion and basic deterrence working as separate pieces on a benchtop rig.
- Built
Rev B · Integrated unit
Edge inference joined to embedded control, human-detect interlock added, safety state machine defined, detection cycle cut from 66 ms to 33 ms.
- In field test
Rev C · Full-scale build
Current model. Full-height structure, closed-loop motion, endstops and homing, expanded safety logic. Open work: long-range targeting accuracy, detection-to-actuation latency, coordinate-to-angle calibration, and behavior under glare, darkness and fog.
- In design
Rev D · Pilot hardware
Sealed weather-resistant enclosure, rangefinder integration, camera and optics upgrade, internal repackaging for serviceability, unit security.
- Planned
Production
Manufacturable design, installation and calibration process, remote monitoring, compliance pathway, support.
Intelligence that moves.
One stack sits under every Ayvion division: perception at distance, precision motion, and a controller that governs what the machine is permitted to do. Wildlife Systems is the first application of it.
Six layers, deliberately separated
Each layer has one job and a defined interface to the next. Perception cannot reach the motors. The safety layer can stop everything above it.
Perception
Planning
Control
Safety
Action
Monitoring
Five things that had to work
Edge inference
Detection runs on an onboard embedded computer, so nothing about the control loop depends on a network round trip. That matters at farms, remote ponds and construction sites where connectivity is weak or absent. The uplink carries status and analytics only.
Detection under real conditions
A real-time detector trained on goose imagery in genuine outdoor light: water glare, backlight, partial occlusion, birds at distance. It runs a human class alongside the goose class, and confidence is a gating input to the safety logic rather than a number on a display.
Mechanics that hold their aim
An azimuth drive carries the assembly; closed-loop steppers move the optical head. Priorities are rigidity, low backlash, reliable homing and repeatable feedback — because pointing error at range is dominated by mechanics, not software.
Camera-to-emitter calibration
Image position becomes a mechanical angle through a maintained geometric relationship between camera, ranging sensor and optical head. Holding that stable through temperature swings, vibration and field service is core to the mechanical design, and refining the process is an active work item.
A controller that can say no
A dedicated microcontroller — not the inference computer — owns motion, homing, endstops, travel limits, fault monitoring, communication-loss behavior and the physical emitter-enable line. If the high-level computer goes quiet, the hardware doesn't coast on its last command; it falls to a safe state on its own.
Always in exactly one known state.
Transitions are explicit and validated. Deterrence is reachable only from tracking, and only when every gating condition holds. A human detection or any fault routes straight to a safe state.
- Startup
- Homing
- Idle
- Searching
- Tracking
- Deterrence
- Safety hold
- Fault
- Emergency stop
Current model
The platform above, applied. Revision C is the full-scale integrated build now in outdoor field testing; Revision D is in design.
| Designation | Goose Deterrent System — Model 1 |
|---|---|
| Revision | C · full-scale integrated build |
| Build status | Outdoor field testing |
| Next revision | D — sealed enclosure, rangefinder, camera and optics upgrade, internal repackaging |
| Perception | High-resolution camera, real-time detection, goose and human classes |
| Compute | Embedded NVIDIA Jetson Orin Nano class |
| Detection cycle | ≈ 33 ms per inference pass |
| Motion control | Dedicated microcontroller, closed-loop NEMA 17 stepper drive, endstops and homing |
| Mechanism | Azimuth rotating drive plus bearing-supported pan / tilt optical head |
| Emitter | 532 nm optical head — heat-sinked, duration-limited, zone-gated, hardware-interlocked |
| Deterrence mode | Patterns traced around the target area — non-contact |
| Ranging | Laser rangefinder integration — in development |
| Enclosure | Weather-resistant housing, sealed cable entry, thermal and condensation management |
| Power | Centralized DC with per-subsystem regulation. Solar and battery in development. |
| Communications | Dedicated local link, acknowledged commands, watchdog on loss |
| Logging | On-device event logs; remote status reporting in development |
Describes Revision C as built. Specifications change between revisions.
The AI can request. It cannot authorize.
Every machine Ayvion builds acts on its own judgment in a place where people are. We design the safety architecture first and let the autonomy work inside it. Below is how that principle is implemented in Wildlife Systems.
A person in frame is a hardware event.
The detection model runs a human class alongside the goose class on every frame. When a person is detected anywhere in view, the safety microcontroller drops the emitter enable line and the platform enters a safety hold. The perception computer is not consulted and no aim command is issued.
This is the interlock we test hardest, because it has to work on the worst day: low light, partial occlusion, someone appearing at the edge of frame.
The deterrent never needs to touch the animal, and it will not operate around people.
Every deterrence event clears all of these
Independent checks, not one function with several conditions. A failure in any of them stops the action.
Human channel clear
No person detected in frame, gated in hardware ahead of any motion or emission command.
Mechanical travel limits
Physical endstops bound the pan and tilt range regardless of what software commands.
Site operating zone
Aiming is restricted to a region defined at installation. Commands outside it are rejected before they reach the motors.
Homing verified
The system will not enter an operating state until it has confirmed a known mechanical reference position.
Independent enable path
The emitter is gated by the safety controller through separate hardware, not by the perception software.
Communications watchdog
Loss of the link between computer and controller forces a safe state rather than continuing on the last command.
Bounded activation
Any deterrence action has a maximum duration enforced in firmware.
Confidence gate
Low-confidence detections do not qualify a target for a deterrence action.
Faults and emergency stop
Hardware faults, invalid transitions and operator stop all resolve the same way: motion halted, emitter disabled.
Engineered as a regulated optical product.
The deterrent is a directional 532 nm optical emitter and we treat it as one. Its containment, interlocks and operating limits are being developed against IEC 60825 and the FDA/CDRH requirements under 21 CFR 1040, alongside state laser law, wildlife regulations including migratory bird rules, and site-specific review.
Commercial deployment means an authorized installation, a defined and enforced operating area, and a site safety review. This is not a consumer device and we don't sell it as one.
Deterrence, not harm.
The system produces a startle response by tracing a light pattern in the area around the birds. It is not aimed at the animal and makes no physical contact. The objective is to move a flock off a property and make it unattractive to return to.
On privacy: imagery is processed on the device, recording is minimized, per-site privacy zones are honored and retention is limited. There is no facial recognition — human detection determines only that a person is present, not who.
We build the machines, not just the models.
Ayvion Industries is an AI and robotics company. We take perception systems out of the datacenter and put them on hardware that has to work at distance, in weather, unattended — and be accountable for what it does.
We build intelligent physical systems that combine computer vision, robotics and embedded control to solve real-world problems autonomously.
Most AI stops at a recommendation. Ours has to move something. That changes the whole engineering problem: a model that is right 96% of the time is interesting in software and unacceptable in hardware unless something independent decides what it's allowed to do with the other 4%.
So we build the full stack — optics, mechanics, firmware, models, power, enclosure — because the failures live in the seams between them, not inside any one part.
How we build
Safety before autonomy
Independent hardware defines what the machine is permitted to do. The model works inside that envelope, and it never gets to widen it.
Edge first
If a system needs a network round trip to make a decision, it doesn't work at a farm, a remote pond or a construction site. Inference runs on the device.
Modular by default
Camera, compute, electronics and mechanics are replaceable subsystems, so the next generation isn't a redesign and a field failure isn't a truck full of spares.
Measured, not demonstrated
Anything can be made to work once for a video. We instrument our systems to log what actually happened, so the claim and the evidence are the same thing.
How the company is organized
Divisions own markets
Each division owns its own customers, hardware, and regulatory surface. Wildlife Systems is the only one with hardware in the field; Site Monitoring and Inspection & Sensing are engineering directions, not products. See the divisions
The platform is shared
Perception, calibration, motion control, the safety controller, outdoor packaging and power are built once and inherited. A division adds the sensing, the response and the compliance work its market needs. See the platform
Nothing ships on the strength of a demo
A division doesn't get a product page until its hardware has survived a season outdoors and produced data we can hand a customer.
What we do in-house
Mechanical & mechatronic
CAD, precision drivetrains, outdoor packaging, rapid prototyping.
Embedded & electrical
Firmware, real-time motion control, safety state machines, power distribution.
Vision & machine learning
Dataset development, model training, edge deployment, tracking logic.
Test & validation
Environmental testing, long-duration reliability, field measurement.
Product development
Architecture, manufacturing direction, service and installation design.
Environmental application
Water quality, wildlife behavior, site assessment, deployment planning.
Where software becomes motion.
The work spans vision, robotics, firmware, controls, mechanical design, optics, power and field deployment — usually in the same week. Current and upcoming areas:
- Mechanical design
- Mechatronics
- Embedded systems
- Computer vision
- Machine learning
- Electrical
- Test engineering
- Internships
Help us prove it on a real property.
We're selecting sites where goose activity is persistent, measurable and expensive. Partners get early access and direct engineering support. We get the field data that makes the product real.
What a pilot involves.
We assess the site
Walk the property, map where the birds are, identify mounting and coverage options.
We install and configure
Hardware mounted, operating zone defined and enforced for your site.
We measure a baseline
Goose activity recorded before the system is armed, so the results mean something.
You provide access and power
A viable mount with sightlines, a grid drop or room for solar, and permission to collect performance data.
You get the review
A written account of what changed at your site: presence, residence time, return frequency, labor.
Tell us about your property
Application received
We'll review your site and follow up by email. If it's urgent, write to pilots@ayvionai.com and reference your location.