Edge AI for defence, space and maritime: AI that works when the link drops

Defence, space and maritime systems share a problem most commercial AI never faces: the network can't be relied on. Links are contested, scheduled or simply absent, and every watt and gram is accounted for. In these domains, running AI on the platform is less an optimisation than a requirement.

What makes these domains different

• Disconnected, intermittent or limited links. A vessel beyond coastal coverage, a satellite between ground-station passes or a platform in a jammed environment has to act on its own.
• Size, weight and power (SWaP). Payload space, battery capacity and cooling are fixed early in a platform's design, and the AI has to fit inside them.
• Scarce bandwidth. Sending raw sensor data home is often impossible, so the platform must decide what's worth sending.
• Assurance. Systems must be tested, certified and trusted by operators before anyone relies on them.

Australian defence policy reflects this. One of Defence's seven Sovereign Defence Industrial Priorities is the development and integration of autonomous systems, and Defence states that it "will continue to invest in the development of autonomous systems across all domains". Another priority covers test and evaluation, certification and systems assurance. The Advanced Strategic Capabilities Accelerator (ASCA) has the mission of accelerating the development and transition of asymmetric capabilities to the Australian Defence Force through innovation.

Defence: sensor fusion on low power

Many defence sensing problems combine more than one kind of sensor. AICRAFT in Adelaide, operating since early 2021, focuses on multi-sensor fusion across electro-optic and radio-frequency data. It designs tailored onboard computers, board support packages and firmware, offers feasibility studies and defence technology consulting, and emphasises peak AI performance at low power consumption.

Other listed firms supply parts of the stack. Xailient runs detection and face recognition inside cameras and names defence among its markets. BrainChip licenses its Akida neuromorphic processor for very low-power vision, audio and sensor processing. Seeing Machines, best known for driver monitoring, also works in aviation with pilot monitoring.

Space: process in orbit, downlink answers

A satellite's downlink is limited by ground-station passes and radio bandwidth, so processing imagery on board can turn a large raw file into a short result. Sydney's Spiral Blue builds onboard AI computers for exactly this. According to its Space Edge product page:

• The SE-1 uses an NVIDIA Jetson Xavier NX rated at 21 TOPS, drawing 3 W idle, 7 W on average and 20 W at peak.
• The SE-2 uses a Jetson Orin NX 16GB rated at up to 157 TOPS (INT8), with a 25 W maximum that can be throttled to 10 W or 15 W.
• Ready-to-run apps include Vessel Detect, Cloud Clipper, Canopy Mapper, Fire Severity and Water Body Mapper.
• The company reports 10 Space Edge computers flown, starting with a Space Edge Zero launched with SatRev in July 2021, and an SE-1 launched aboard a Satellogic satellite in January 2023.

Space adds problems the ground doesn't have. There's no air to carry heat away, so thermal design relies on conduction and radiation, and the electronics must tolerate radiation. Power is shared with the payload and radios, which is why configurable power modes matter as much as peak TOPS.

Maritime: perception and autonomy at sea

At sea, glare, spray, swell and a moving horizon make visual detection hard, and a vessel may be days from port. Fremantle's Greenroom Robotics builds software for this environment. Its GAMA system is described as a system-agnostic turnkey solution to marine autonomy, remote operations and enhanced crews, for vessels up to 100 m. Lookout+ provides vision-based AI for detecting marine features, hazards, vessels and items of interest, and MIS-SIM is a simulation environment for mission rehearsal, monitoring and playback.

Greenroom says GAMA powers Fremantle 02, an optionally crewed survey vessel that will support Royal Australian Navy hydrographic survey work, and the company has teamed with WA shipbuilder Austal on maritime autonomy.

Engineering for disconnected operation

Whatever the domain, systems that must work without a link share some design habits:

• Degrade gracefully. Define what the system does when a sensor fails or the model is uncertain, and make that behaviour safe.
• Store and forward. Keep results, logs and selected raw data locally, then send them in priority order when a link opens.
• Small, signed updates. Model and software updates may travel over thin links, so keep them compact and verify them on the device before installing.
• Keep people in control. The Australian Government's Guidance for AI adoption recommends clear intervention points where people can pause, override, roll back or shut down AI systems.
• Know your versions. Every result should be traceable to the model, data and software that produced it.

Testing before deployment

Field trials at sea, in the air or in orbit are expensive and sometimes impossible to repeat. Simulation and hardware-in-the-loop testing fill the gap: the real computer runs the real model against simulated sensors and scenarios, including rare and dangerous ones. Greenroom's MIS-SIM links real and synthetic worlds for rehearsal and playback, and Vedametric trains its robot controllers in simulation and then moves them onto physical machines. The evaluation discipline is the same as for any edge model: an agreed test set, measured on the target hardware, with results recorded for assurance. The compression guide covers how to keep that discipline while shrinking a model.

Questions to settle before engaging a supplier

• Offline behaviour. Can the full system, including the model, run with no external connection, and for how long?
• Power modes. What accuracy and throughput does the system deliver at each power setting, measured on the target hardware?
• Supply chain. Where do the boards, modules and key components come from, and what is their expected production life?
• Update and rollback. How are models updated in the field, and how is a bad update reversed?
• Evidence. What test data, simulation results and trial records will be handed over for assurance?
• Export and ownership. Which export controls might apply to the hardware or software, and who owns the resulting IP?

Firms at a glance

FAQ

What does SWaP mean for edge AI?

SWaP stands for size, weight and power. On aircraft, satellites, vessels and portable equipment these limits are set early in the design, so the AI hardware and model must fit within them. That is why low-power accelerators and configurable power modes matter so much in these domains.

Can satellites run AI models on board?

Yes. Spiral Blue reports 10 of its Space Edge onboard AI computers flown. Its SE-2 uses an NVIDIA Jetson Orin NX module and runs apps such as vessel detection, so the satellite can downlink results instead of raw imagery.

How is AI for vessels and aircraft tested before deployment?

Largely in simulation and hardware-in-the-loop set-ups, where the real computer runs the real model against simulated sensors, followed by staged field trials. Greenroom Robotics, for example, offers MIS-SIM for mission rehearsal, monitoring and playback.

Where does autonomy sit in Australian defence industry policy?

Development and integration of autonomous systems is one of Defence's seven Sovereign Defence Industrial Priorities, which describe the industrial capabilities Defence requires in Australia.