Adelaide artificial-intelligence venture Metacognition has assembled A$15 million of pre-seed funding around an ambitious proposition: robots and industrial agents will need persistent memory, controlled reasoning and the ability to act over long periods before they can become dependable colleagues.
The company was founded in 2026 by professors Anton van den Hengel and Stephen Gould with research strategist Dr Paul Dalby. Main Sequence committed the initial A$10 million, followed by a A$5 million investment from Airtree announced in late October.
That is a large amount of capital for a company at this stage. It reflects the founders’ research and commercialisation records as much as the maturity of the product. Metacognition has released code and benchmark results, but it has not disclosed revenue, signed enterprise customers or commercial deployments.
Building beyond a short AI conversation
Metacognition begins with a limitation familiar to buyers of generative AI. Large language models can reason within a prompt and retrieve external information, yet long-running systems still struggle to preserve reliable memory, manage competing goals and respond safely when the physical environment changes.
That becomes more consequential when software controls a robot, a mine workflow or an electricity system. An inaccurate answer on a screen can be corrected. A machine acting on stale assumptions can interrupt production, damage equipment or create a safety risk.
The startup is developing ZEOS, which it describes as an operating system for agents and robots. The public code separates tasks, schedules priorities and controls how model-generated instructions can trigger effects. A higher-priority event can interrupt a lower-priority job, while protected boundaries are intended to stop untrusted text from directly authorising actions.
The commercial ambition is to sit below language models and above machines and enterprise systems. Instead of selling a single robot, Metacognition wants to provide the orchestration and memory layer that helps different hardware and models work together.
A founding team shaped by research and commercialisation
Van den Hengel is chief scientist of the Australian Institute for Machine Learning and director of the Centre for Augmented Reasoning at Adelaide University. He founded the institute and later worked as a director of machine learning at Amazon, giving him experience across public research, large technology platforms and commercial applications.
Gould has combined academic computer-vision research with startup and industry work. Metacognition says he was a principal research scientist at Amazon, a director of the Australian Centre of Excellence in Robotic Vision and a co-founder of Sensory Networks, which Intel acquired in 2013.
Dalby’s background is in research strategy, partnerships and institutional funding. That matters because the company is attempting to turn a broad research architecture into a product that industrial buyers can evaluate, govern and support.
The founders’ credentials help explain the early investor support, but they do not eliminate the transition risk between a strong research team and a repeatable software business.
NavGPT-3 provides an early proof point
Metacognition has released NavGPT-3, an open-source vision-language model designed to translate verbal instructions into robot navigation. The company says it outperformed models from Alibaba and Mistral in simulated benchmarks. Those results are encouraging, but they remain company-reported tests rather than evidence of sustained performance in mines, warehouses or other live environments.
The distinction is important. Navigation benchmarks measure whether a model reaches a destination under defined conditions. Industrial adoption also requires cyber security, audit trails, integration with existing control systems, reliable exception handling and clear allocation of responsibility when an automated decision is wrong.
ZEOS is meant to address part of that gap by making actions more structured and inspectable. Its public repository is licensed under the GNU Affero General Public License, which can encourage external testing and contribution. Metacognition will still need to explain how an open core converts into paid deployments, support contracts or enterprise licensing.
The buyer and business-model test
The initial buyer is likely to be an organisation running complex, always-on operations rather than a consumer purchasing a household robot. Mining, energy, logistics and advanced manufacturing offer environments where improved autonomy has a measurable value, but they also impose strict safety and integration requirements.
Metacognition has demonstrated a simulated open-cut mine in which an operator directs activity through natural language and the system responds to changing ore grades or stop instructions. A simulation can show the architecture, but a paid pilot will need defined scope, accountable human oversight and proof that the system creates value without weakening operational controls.
The company will compete with robotics platforms, industrial-automation suppliers, cloud providers and internal engineering teams. Large vendors already control many customer relationships and could add memory or orchestration functions to their existing products.
Procurement will also be slower than the pace of the underlying research. Industrial customers will expect security review, integration testing, support commitments and evidence that a startup can maintain the platform for the life of an asset. Converting research interest into contracted recurring revenue will therefore require patient enterprise sales and implementation capability alongside the core engineering team.
Metacognition’s defence is that its architecture is being designed around long-term reasoning and controlled action rather than adding those features to a conversational model. Whether that distinction survives contact with rapidly improving foundation models is a central commercial question.
What A$15 million buys
The funding gives the company capacity to recruit a highly technical team, develop the architecture and place prototypes into demanding environments. Van den Hengel has said the business expects to hire about 20 people, many with doctoral-level training, and recognises that a much larger research organisation would require further capital.
That creates a disciplined sequence of proof points. Metacognition must turn research claims into a stable developer platform, secure credible pilot partners, demonstrate safe performance and identify the product boundary customers will pay for. It must do so before research-intensive hiring consumes the early funding.
Australia has produced strong AI researchers but often captures less of the downstream platform value. Metacognition is explicitly trying to keep foundational capability and commercialisation together in Adelaide.
The A$15 million round gives that thesis a serious start. The next chapter will be written by live deployments: systems that remember correctly, pause when they should and create enough operating value for customers to make ZEOS part of their long-term technology stack.
Featured image: Igor Omilaev / Unsplash. Conceptual AI and robotics illustration; it does not depict Metacognition products or staff.
Sources: Metacognition company and founder information; Metacognition funding announcement; Airtree portfolio record; ZEOS public repository and licence; Adelaide University profile of Anton van den Hengel; Forbes Australia, 28 October 2026; Capital Brief, 10 September 2026.
