The aircraft captures everything. The AI decides what gets reviewed first.
DJI's 2026 Onboard AI Challenge winners show a shift from “fly first, inspect everything later” toward a mission where detections, locations and priority findings can surface during the flight.
DJI's 2026 winners show how enterprise drones can perform the first analytical pass during the mission. They can detect an object, classify it, attach a location, rank urgency and surface an alert sooner. A deployable Canadian AI workflow still needs representative data, validation, suitable compute, software integration, human review, operating procedures and the aviation approval required for the mission.
DJI rewarded practical aerial AI — not abstract model benchmarks.
DJI named the winners of its Enterprise Drone Onboard AI Challenge 2026 after a global competition focused on practical field applications.
Eligible solutions were built for the Matrice 4 Series, Matrice 4D Series, DJI Dock 3, Matrice 400 or Manifold 3. DJI's summary highlighted crop counting, bridge-crack screening, pollution and litter detection, search-and-rescue support and transportation inspection.
The common thread is not “AI flies the mission alone.” It is that imagery can be interpreted earlier, reducing the time between capture and the operator's first useful review.
Best Onboard AI Model award winners named by DJI.
Industry Application Excellence award winners.
Supported families: Matrice 4, Matrice 4D, Dock 3, Matrice 400 and Manifold 3.
Move the first review from the desk into the mission.
Onboard AI changes where the first triage happens. It does not remove the reviewer.
Everything waits in the same queue.
Every frame can enter the same review queue, even when most imagery contains no priority finding.
Flagged evidence moves forward first.
Operators can review likely findings sooner while preserving source evidence and a human-confirmation path.
Five winning application patterns worth watching.
These are workflow examples. An award or field test does not guarantee the same model will perform identically at another site, season or operating condition.
Agriculture
AgroCount AI paired a planned Matrice 4E + Manifold 3 architecture with onboard processing and geotagged plant counting at a commercial banana plantation.
Transportation
A multi-function project combined inspection and enforcement functions across FlightHub 2 and Manifold 3.
Infrastructure
Visual screening can flag likely cracks or anomalies while preserving location and source imagery for qualified follow-up.
Environment
Pollution and litter detections can direct sampling, cleanup or verification resources toward mapped findings.
Public safety
AI may flag possible people, vehicles or heat signatures for rapid review, where missed detections and false positives can have serious consequences.
The value is a shorter path from capture to triage.
What matters after detection is different for utilities, public safety, municipalities and inspection teams.
A utility may want anomalies ranked before landing. A search team may want an alert during the mission. A municipality may want recurring detections written into a work-order system. The same aircraft can support very different integration projects.
Priority findings can reach an operator sooner than a full post-flight review.
Teams may transmit detections and selected evidence first instead of every frame.
Analysts can start with flagged items but still need quality-control sampling of unflagged data.
A fixed model can make the first pass more consistent only within validated conditions.
Value grows when detections enter the customer's map, report, dispatch, maintenance or asset workflow automatically.
Buying AI-capable hardware is not the same as buying a finished AI solution.
The aircraft and compute are the platform. A deployable application needs evidence, integration and operating controls around them.
Task definition
Define exactly what the model must detect, count or classify and what decision follows.
Training data
Use imagery representing target, camera, altitude, season, weather and site variation.
Model validation
Measure misses, false alerts and performance limits on held-out operational data.
Onboard compute
Confirm model format, processing load, hardware compatibility and thermal limits.
Integration
Route results into FlightHub 2, GIS, dispatch, inspection or asset systems.
Human review
Assign who confirms findings and how uncertain or unflagged evidence is checked.
Operating procedures
Document versions, pre-flight checks, alert handling, escalation and fallback.
Change control
Retest after camera, firmware, model, route, season or site conditions change.
Choose the operating model first. Then choose the DJI Enterprise platform.
Portable field deployment, modular onboard compute and repeated dock missions create different hardware and integration requirements.
Matrice 4 Series
Matrice 4E fits compact mapping and visual-data workflows. Matrice 4T adds thermal sensing for inspection, public safety and lower-light missions.
Matrice 400
Matrice 400 supports Manifold 3 and broader payload integration for demanding inspection, public-safety, geospatial and custom-development programs.
DJI Dock 3
Dock 3 supports recurring Matrice 4D Series and FlightHub 2 workflows. Ownership does not itself authorize BVLOS or remote operations in Canada.
View DJI Dock 3Manifold 3 is the compute layer — not the business outcome.
DJI rates Manifold 3 at up to 100 TOPS and lists support for Matrice 4, Matrice 4D and Matrice 400. Compute capacity still needs a model, deployment environment, integration plan and validation evidence.
Start with the decision the AI must support.
Ask the vendor or developer to separate what works now, what requires custom development and what evidence will be delivered at acceptance.
A polished demo video is not a substitute for a representative site test.
A narrow task is easier to validate than “inspect the site with AI.”
Ask whether Canadian winter, low light, haze, vegetation and site materials are represented.
Set thresholds for misses and false alerts, then define the human review plan.
Aircraft, Manifold 3, edge device and cloud inference create different latency and security trade-offs.
Confirm output format, geotag, evidence, alert route, API and reporting responsibility.
Define version control, retraining triggers, firmware compatibility and response time.
Preserve safe flight and a manual review path when model, network or integration is unavailable.
AI capability and Canadian flight authority are separate questions.
What the technology can support
Detection, counting, geotagging, first-pass screening, alerts and integration with a managed enterprise workflow.
What the operator must still establish
Operating category, pilot and operator qualifications, aircraft safety-assurance status, airspace access, privacy, cybersecurity, site procedures and human accountability.
Transport Canada's Level 1 Complex framework permits specific lower-risk BVLOS operations only when its operating, pilot, operator-certificate, airspace and aircraft requirements are met. Other missions may need another authorization. An onboard AI model does not change those requirements by itself.
DJI onboard AI questions
What is DJI onboard AI?
DJI uses the term for AI algorithms that can run on supported enterprise drones or the Manifold 3 computing platform. The model can analyze imagery closer to the flight and return detections, counts or alerts without waiting for a complete post-flight workflow.
Which platforms supported the DJI Onboard AI Challenge 2026?
DJI listed Matrice 4 Series, Matrice 4D Series, DJI Dock 3 with Matrice 4D Series, Matrice 400 and Manifold 3 as eligible platforms.
Does Matrice 4E include the AgroCount AI solution?
No. AgroCount AI was a developer project with a planned Matrice 4E and Manifold 3 architecture. Buying a Matrice 4E does not automatically include that model, its training data, integration or field-validation work.
Can onboard AI replace a human inspector?
It should be treated as a screening and prioritization tool unless the customer has validated a broader role. Qualified people still need to confirm findings, review uncertain results and remain accountable for engineering, emergency-response or regulatory decisions.
Does onboard AI work without an internet connection?
Some inference can run locally on supported onboard hardware, but the complete workflow may still use a link for alerts, maps, remote supervision, software integration or data transfer. The answer depends on the selected architecture.
Is Manifold 3 required for every DJI onboard AI application?
No. DJI's challenge included drone deployment, Manifold 3 deployment and cloud-based deployment architectures. The compute choice depends on the model, supported aircraft, latency, interfaces and operating workflow.
Does AI authorize automated or BVLOS drone flights in Canada?
No. The operator must meet the Transport Canada requirements for the intended mission. The aircraft, pilot qualifications, operator certificate, airspace, procedures and any required authorization remain separate from the AI function.
Can SpeedyDrone help plan an enterprise AI drone workflow?
SpeedyDrone can help compare current DJI Enterprise aircraft, Dock 3, Manifold 3 compatibility, payloads, controllers, software and deployment requirements. Custom AI development, validation, integration, privacy and regulatory work must be scoped to the customer's exact project.
Sources and current product records
- DJI: Onboard AI Challenge 2026 winners announcement
- DJI Developer: official challenge site and rules
- DJI Enterprise: Manifold 3 specifications
- DJI Enterprise: FlightHub 2 third-party algorithm support
- Transport Canada: Level 1 Complex operations
- SpeedyDrone: DJI Matrice 4E
- SpeedyDrone: DJI Matrice 4T
- SpeedyDrone: Matrice 400 SP Plus Combo
- SpeedyDrone: Matrice 400 SP Plus Full Package
- SpeedyDrone: DJI Dock 3
Sources and current SpeedyDrone product records checked August 22, 2026. Challenge results describe submitted projects and DJI's published summaries. They are not performance guarantees for a Canadian deployment. Specifications, firmware, software support, product availability and regulations can change.
Build the detection workflow before you buy the AI hardware.
Send SpeedyDrone the target object or condition, site, aircraft preference, mission frequency, required alert time, current software, data-security requirements and deployment date. We can help separate hardware from the development and integration work that still needs to be scoped.
