Why we are teaching robots what language models can already do

Language models can plan, understand instructions and use tools. Robots can move and grasp precisely. What is missing is the link between the two, and it has to work reliably in everyday life.

That is where we come in. Starting today, we are applying our experience from AI projects to robotics.

The problem: staff are missing where people are needed

Hotels and elderly care are short of skilled staff. A large share of their time goes into errands, inspection rounds and tidying up, which are tasks that do not need human attention. That time is then missing for the people they care for.

Our approach

We are not building a robot for a single task. We are building a robot that learns new tasks through language and a growing library of skills:

• A language model plans: It breaks an instruction such as "Bring the towels to room 204" into individual steps.
• Tested skills execute: Driving, grasping, handing over, reporting.
• Simulation first: Every skill is tried out thousands of times virtually before it runs on real hardware.

What comes next

In the new year we will build our simulation and development environment and bring the first skills to real robot platforms. We will report here regularly on what works and what does not.