CSIRO launches Vetra edge AI hub for robots in Queensland

CSIRO has launched its Vetra AI infrastructure in Queensland, designed to process AI data on site for robots and sensing systems.

The installation is based at the Queensland Centre for Advanced Technologies in Pullenvale, alongside what CSIRO describes as Australia's largest robotics research facility. It is intended to let machines process information closer to where it is generated, rather than relying on distant cloud systems.

That approach, often described as edge AI, is attracting wider attention as artificial intelligence moves into machines, sensors and industrial systems that must react quickly in physical environments. In those settings, delays caused by sending data to remote data centres can be a constraint, particularly in safety-critical applications.

Vetra forms part of an "edge-core-cloud" model with CSIRO's larger supercomputing systems in Canberra. Under that arrangement, immediate processing happens locally, while more intensive analysis can be handled later by larger computing centres.

Vetra includes 48 graphics processing units, or GPUs, which are commonly used for AI workloads because they can perform many calculations in parallel. The infrastructure is also modular, allowing it to be expanded if demand increases.

Liming Zhu, director of Data61, said the project reflects a shift in how AI infrastructure is being deployed.

"AI is rapidly moving beyond digital systems into the physical world, including robots, infrastructure, sensing and safety-critical environments," Zhu said.

"Vetra enables real-time physical AI research by bringing high-performance computing to the edge, where proximity to data allows systems to respond, learn and operate safely in complex environments in ways that are not possible with cloud-only or distant data centre approaches."

He described the move as part of a broader discussion about sovereign AI and the location of critical digital infrastructure.

"This represents a different form of sovereign AI, where physical location becomes part of the capability itself, establishing a model and associated innovative technologies that can be replicated and exported to other locations where on-site, trusted AI is required."

Sustainability focus

One of the main engineering challenges in AI computing is cooling. Systems running dense AI workloads generate large amounts of heat, and conventional cooling methods can consume significant amounts of water and energy.

Vetra uses a carbon dioxide-based cooling system intended to reduce reliance on more water-intensive methods. Under normal operation, CSIRO said, the system wastes almost no water for cooling and could save about 225 tonnes of carbon dioxide emissions a year, roughly equivalent to removing 50 cars from Queensland roads.

Angus Macoustra, chief technology officer at CSIRO, said those considerations were built into the project from the start.

"High-performance AI systems generate a lot of heat in dense, enclosed spaces. Vetra shows how advanced technology can be delivered in a way that significantly reduces water use and emissions," Macoustra said.

Industry support

The project was delivered with support from Australian small and medium-sized businesses including Oper8 Global and XENON, alongside unnamed global technology partners. CSIRO said the companies contributed to the design, delivery and installation of the computing environment.

Mike Andrea, chief technology officer at Oper8 Global, outlined the engineering requirements for the site.

"Our role was to design and deliver a highly specialised, modular computing environment that could support extremely dense AI workloads in a compact footprint. Infrastructure like Vetra requires a very different approach to traditional data centres, particularly when integrating advanced cooling systems, high-power GPU infrastructure, and strict reliability and safety requirements," Andrea said.

He said the project's physical design was shaped by its proximity to robotics laboratories.

"Working with CSIRO meant translating cutting-edge research needs into physical infrastructure that can operate day in, day out, close to robotics laboratories and real-world test environments. It's a strong example of how Australian engineering and construction expertise can support next-generation digital infrastructure," Andrea said.

XENON was responsible for the design, installation, testing and commissioning of Vetra's high-performance computing systems, networking and software. Dragan Dimitrovici, chief executive officer and founder of XENON Systems, said the system was designed to align closely with researchers' needs.

"XENON's focus has been to help the CSIRO team bring cutting-edge AI infrastructure to the edge, where researchers work and physical AI is taking shape. From the initial concept stages through to delivery in Pullenvale, XENON has worked closely with CSIRO to ensure the infrastructure matches researchers' requirements and enables the next generation of AI innovation.

"XENON is proud of its long-standing partnership with CSIRO and looks forward to providing ongoing support to ensure these valuable computing resources remain highly available and are used to their full potential by CSIRO researchers. Being part of this project demonstrates how Australian AI infrastructure expertise and engineering capability can deliver trusted, mission-critical sovereign AI infrastructure for research and industry," Dimitrovici said.

Within CSIRO, the system is also tied to research into embodied AI, in which machines learn through direct interaction with physical environments rather than only from simulations or internet-based datasets. Peyman Moghadam, head of CSIRO's Embodied AI Cluster, said the missing link had been local processing close to the machines themselves.

"Robots and physical AI systems need to keep learning from the physical world, not just from internet datasets or simulations," Moghadam said.

"Vetra gives us the missing edge layer for this workflow, helping turn real-world robotics data into better, safer and more adaptable AI systems."