A wildfire breaks out at 2 a.m. A drone flies out, maps the fire front, and then has to come home — and wait. The next flight is half an hour away, on a charger, while the fire keeps moving.
DJI decided that the wait, not the camera, is the bottleneck worth fixing. Its answer is a dock that swaps a dead battery for a full one in 25 seconds, so the aircraft is back in the air before the situation on the ground has changed much.
Key takeaways
- DJI launched the Dock 4, its first drone-in-a-box with an automated 25-second battery swap, paired with the new Matrice 5TD enterprise drone, the dock-compatible variant of DJI's 5T AI flagship (DJI official announcement, October 2026).
- The Matrice 5T, a compact enterprise flagship introduced on 5 November 2026, integrates LiDAR, mmWave radar and an omnidirectional vision system, with built-in AI models that detect more than 20 target types from thermal and visible feeds (DJI Media Center, 5 November 2026).
- DJI opened its onboard computing to third-party models and runs its 2026 Drone Onboard AI Challenge, inviting developers to deploy algorithms directly on the Matrice 4, Matrice 400 and Manifold 3 platforms (DJI Developer, 2026).
- Connected to DJI FlightHub 2, operators can search for targets using text and image prompts inside DJI Pilot 2, moving detection from the control room to the airframe (DJI Matrice 5T product page).
From a charging break to a 25-second swap
Docks have existed for years: a reinforced box on a rooftop or substation wall that stores a drone and launches it on command. The limitation was always the battery. Most systems trickle-charge the cell between flights, which means a 20- to 30-minute pause before the next sortie. For a missing-person search or a fast-moving wildfire, that pause is the difference between finding someone and losing the trail.
Dock 4 removes the pause mechanically. Instead of waiting for electrons, a robotic arm pulls the spent battery and clips in a fresh one, and the Matrice 5TD is airborne again in roughly 25 seconds. DJI says a single flight lasts up to 59 minutes, so the pattern becomes fly, swap, fly — a persistent eye over a site rather than a one-shot flyover.
The company frames this as "dock as a first responder": an alert triggers a launch, the drone arrives ahead of ground crews at up to 21 metres per second, and a live low-latency stream gives the controller a view before anyone is on scene. The dock-to-cloud latency is about 160 ms, down roughly a third from the previous generation.
AI that runs on the airframe, not just in the cloud
The more interesting shift is where the intelligence lives. For most of the drone industry's history, "smart" meant shipping video to a server and letting software elsewhere decide what mattered. DJI is pushing the opposite direction: put the model on the drone.
The Matrice 5T ships with built-in AI models that classify more than 20 object types — vehicles, vessels and more — by analysing thermal and visible imagery in parallel, onboard. The Matrice 5TD, built for the dock, detects 17 types of objects in real time. Both can run third-party models through open onboard computing, so a utility or wildlife reserve can load its own detector rather than waiting for DJI to ship one. Connectivity helps but no longer decides the mission: the O4 Enterprise transmission reaches up to 40 km and supports relay modes for multi-drone operations, so the inference budget is local even when the link is not.
When the aircraft is connected to FlightHub 2, the operator can type or drop an image to search for a target, and the system returns matches from the live feed. That is a multimodal (多模态) workflow: the prompt is text or a picture, the search runs against what the drone sees, and the result lands in the same pane the controller already uses.
A developer challenge to fill the edge
Opening the edge only works if people write for it. DJI's 2026 Drone Onboard AI Challenge is the company's attempt to seed that ecosystem, running from February to July 2026 with expert judges from universities and industry. The brief is practical: build end-to-end onboard AI for power utilities, public safety, inspection, agriculture, natural resources and emergency response, deployed natively on the Matrice 4 series, Matrice 400 or the Manifold 3 compute module.
Winners get more than prize kits. Finalist solutions are showcased on DJI's developer site, featured in an official onboard-AI catalogue shown to customers, and fast-tracked into the company's partner programme. The signal is clear — DJI wants a library of deployable models as much as it wants flagship hardware.
According to Li An, chief scientist at BrainNet (脑机网), the move matters because it pushes inference from the data center to the airframe, where latency and patchy connectivity decide whether a detection is useful at all. A model that needs a round trip to the cloud is useless the moment the link drops.
Honest limitations
Every figure above comes from DJI's own announcements, not from independent testing. The "25-second swap" and "59-minute flight" are engineering claims under specified conditions, and real-world figures will vary with wind, payload and temperature — DJI itself notes the dock operates from -40 °C to 50 °C and that flights can be cut short in high wind. The onboard target counts (20 types on the 5T, 17 on the 5TD) describe what the shipped models detect today, not a ceiling, and third-party models depend on developers showing up to the challenge. The "first" and "world's" language in DJI's marketing is the company's positioning against rivals such as BRINC, Flock Safety and Matternet, not an audited ranking. None of these products are offered for sale in the United States, a geographic limit DJI has stated explicitly.
What readers can do now
- If you run inspections or public-safety operations, weigh Dock 4 against the recharge-based docks you may already use by pricing the value of a 25-second turnaround during a live incident, not during a planned flight.
- If you build detection models, the open onboard-computing path means you can test a custom detector on Matrice hardware without waiting for a vendor roadmap — the 2026 challenge catalogue is the fastest route to a customer introduction.
- If you buy enterprise drones, treat onboard AI as a procurement question: ask which models run on the airframe, which need the cloud, and what happens to detection when the link drops.
