AI-RAN: The brain in the backpack

Imagine carrying your brain in a backpack. The smarter you get, the heavier it becomes. At some point, you have to choose between thinking and moving.  

This is the dilemma robots face today, and humanoids most of all. The AI models that help them understand tasks, plan movements and adapt to their surroundings need a lot of computing power. On board, that power adds weight, drains the battery and produces heat. Every watt spent on thinking is a watt missing for moving, lifting and gripping. 

The way out: let part of the brain stay behind. This is the idea behind distributed inference. Reflexes stay on the robot. Safety functions and motion control react instantly, without waiting for the network. The heavy thinking, such as planning and predicting movements, moves to powerful GPUs nearby. Training the models continues in large AI factories.  

For this to work, the GPUs have to be close enough to respond quickly and available wherever robots operate. This is where AI-RAN comes in.

The radio access network, or RAN, is the part of a mobile network that connects devices to it. It consists of antennas, base stations and the computers that process the radio signals.    

This works through distributed AI infrastructure, which brings compute capabilities closer to users and devices by integrating them into the telco network edge, at locations between the central data center and the RAN, such as aggregation sites or core sites.  

This in turn enables third-party AI workloads to run inference closer to where data is generated and consumed, reducing reliance on centralized data centers. The result is one network that offers connectivity and AI compute, distributed across many locations and much closer to robots than a central data center. The backpack gets lighter, and the robot still has access to all that brainpower. 

Together with Deutsche Telekom, Nokia and NVIDIA, we have tested this approach in a lab proof of concept. Deutsche Telekom runs NVIDIA AI Aerial with NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs at the edge of its network. Our dual-arm Franka Research 3 setup draws on this computer to perform individual tasks in server assembly.  

The partners validated that an Agile Robots AI workload can be orchestrated between the Industrial AI Cloud and the telco network edge. Inference at the edge supports the robot's movement planning and prediction.  

This is an early step in a lab setting, but it shows that the architecture can work. 

From here, we are looking in two directions.  

Stationary industrial robots could use distributed compute to run larger AI models than a single cell can handle. Updates could reach a whole fleet at once, and the hardware on each robot could become leaner.   

For mobile robots such as humanoids, the backpack matters even more. On the factory floor and beyond, they could stay lighter and run longer while drawing on intelligence from the network. We are working on capabilities that would benefit from edge compute. Robots could share what they learn across a fleet.   

They could reason through a task step by step before acting. With world action models, they could even “dream” through their next moves before making them and spot possible collisions earlier. Today, these are research directions. AI-RAN could help turn them into reality. 

The brain no longer has to fit in the backpack, and that could change how robots are built.  

Until now, how smart a mobile robot can be has depended on how much computing power it can carry. If the network supplies the intelligence, compute could stop being the limiting factor for mobility. Manufacturers could design robots around the job instead of around the processor: lighter, with longer battery life and leaner hardware, while their capabilities keep growing through the network. For telecom operators, the mobile network becomes an infrastructure for physical AI.  
With Deutsche Telekom, Nokia and NVIDIA, we are moving the heavy thinking into the network, so that robots can get smarter without getting heavier. 

Header image: A humanoid robot walking alongside a person. For illustrative purposes only.