The robot of the future might not be programmed line by line. We could show it a task, correct it, and explain verbally what we want, just as we would with a new colleague. That is the premise of Metacognition AI, a newly formed Australian lab that has raised 10 million Australian dollars from Main Sequence, the deeptech fund backed by CSIRO, in a pre-seed round roughly ten times the Australian average for this stage, according to Forbes Australia.
The company was founded by Anton van den Hengel, chief scientist at the Australian Institute for Machine Learning, alongside Stephen Gould and Paul Dalby. It does not intend to compete directly with OpenAI or Anthropic in building the largest foundational models. Instead, it aims to build what it describes as an operating system to use those models in a more controllable, personalized, and secure way.
The model is not the complete product
A large model can recognize images, understand instructions, and plan actions, but turning these capabilities into a reliable agent requires memory, control, user adaptation, and integration with physical or digital tools. Today, this is a layer that every company tends to build on its own.
Metacognition is betting that this layer can become a common platform, much like an operating system abstracts away hardware complexity for application developers.
For robots, programming every exception does not scale
A traditional factory works well when the environment and tasks are highly predictable. But a robot that must work alongside people or handle diverse objects constantly encounters situations that were not coded in advance.
Learning through language and demonstration promises to lower reconfiguration costs. An operator who knows the job could teach the machine directly without having to become a robotics expert.
Voice is an interface, not the real innovation
Telling a robot "pick up that box and place it on the pallet" is the visible part. Behind it, there must be perception, planning, motor control, error handling, and a model of what the user intends. The operating system's value lies in coordinating these components while maintaining safety constraints.
This is especially critical when AI operates in the physical world: an incorrect response on a screen can be deleted, but a wrong movement by a machine cannot.
Australia tries to compete without building a frontier model
For ecosystems with less capital than the United States, trying to replicate labs that spend billions on compute is difficult. Metacognition is choosing a different spot in the value chain: utilizing existing models and building technology around their deployment.
It is a potentially more defensible strategy if the platform builds expertise in interaction, safety, and adaptation in real-world environments. Main Sequence, founded to commercialize Australian scientific research, is backing precisely this possibility.
The real test will be outside the lab
The "Windows for robots" metaphor is powerful, but it also sets a massive target. Windows worked because relatively stable hardware standards existed; robotics is a far more fragmented landscape of sensors, actuators, form factors, and requirements.
To truly become a common layer, Metacognition will have to prove that it can generalize across different platforms and deliver enough reliability to be deployed in production environments. Ten million funds the start of this test.
If the company succeeds, the most compelling shift will not be having outright smarter robots, but making robotic intelligence accessible to the people who actually know the job. Programming could become less and less like writing code, and more and more like teaching.
