The Tree of Robots: A living encyclopaedia for intelligent machines

For decades, robots have been classified based on their form and underlying technology – what they look like they should do, as well as the narrow tasks they have been specifically designed to perform.  

But what if we were to dig deeper and categorize them not by appearance or purpose, but by their abilities? And not just those abilities that are overt and obvious, but those that might be hidden, or ancillary to the main function? 

This is exactly what Professor Sami Haddadin, Vice President of Research at MBZUAI and his Ph.D. students from TUM (Technical University of Munich) are doing with the Tree of Robots — a groundbreaking new framework that fundamentally rethinks how we understand, compare and develop robotic systems, with potentially profound implications. 

“This is the first systematic attempt to create a framework to fundamentally categorize any robot, machine, or embodied AI system in terms of their morphology and capabilities rather than in terms of descriptive objectives,” says Haddadin, who likens the Tree of Robots to Charles Darwin’s ambitious Tree of Life; a model that sought to illustrate how species are related to each other throughout evolutionary history. 

“Rather than saying ‘I have developed a technology, let’s see what to do with it’, the Tree of Robots provides a systemic, fundamental understanding of the desired capabilities and actual abilities of a system, while what physical technology they are rooted in is captured in their morphology. It’s taking a fundamental paradigm shift and perspective change into what an intelligent machine truly is, especially when considering the fact that AI and control allow robots to advance by programming alone.” 

Haddadin’s work is presented in the paper Kirschner, R.J., Karacan, K., Melone, A. et al. Categorizing robots by performance fitness into the tree of robots, Nat Mach Intell (2025). The research was selected to be on the front cover of the March, 2025 issue of renowned journal Nature Machine Intelligence.

With Haddadin and Robin Kirschner being the lead authors, the paper details the workings of the new framework, which follows a structured, reproducible procedure using its Process Database and Metrics Definitions to obtain process-based fitness metrics, as well as a final fitness score (i.e. how effectively and efficiently a robot can complete a specific task), and classification within the Tree of Robots. 

Kirschner details this core concept for the industrial space: “The Process Database in the industrial robotics category offers detailed sections such as ‘handling’ – which includes subcategories like ‘tool exchange,’ ‘machine steering,’ and ‘pick and place’ – as well as ‘assembly and disassembly,’ covering tasks like ‘wiring,’ ‘levering,’ and ‘screwdriving.’ Additionally, the Metrics Definitions section categorizes key performance indicators, including ‘force sensing,’ ‘manual maneuverability,’ and ‘human safety,’ with specific metrics like ‘guiding force,’ ‘minimum motion force,’ and ‘maneuver effort’ providing critical insights into a robot’s fundamental capabilities.” 

Putting robots through this procedure helps us to ascertain and understand their ability to perform particular movements and physical interactions with the environment and humans, and which designs work best in certain situations: “The basis of this ambitious attempt” says Kirschner “is to compile the definitive guide to today’s and future robots and their full capabilities.” 

Beyond classification 

Laying the foundations for such a compendium is an achievement in itself, but while the Tree of Robots provides a powerful model for classification, it has the capacity do so much more.  

As well as helping researchers and developers understand which robotic designs work best in certain situations, and thus guide future improvements, this method has the potential to discover robots’ ‘hidden’ capabilities — giving us a much better understanding of what robots are truly capable of and what they can deliver at scale. 

“A computer is a computer, but we don’t see the program of it,” says Haddadin. “We use it for certain things, but we don’t necessarily know what more it can do with a different program, let alone with an ever changing and learning AI algorithm running on it. 

“And to take it further, look at humans. When we have our online conversation, I see you but the only thing I learn form that interaction is that you are able to speak and respond by formulating questions on a particular subject. I cannot tell anything else about you. Are you an athlete? Are you able to perform or even compose music? Do you do poetry? What other capabilities and skills do you have?” 

Uncovering the full extent of a robot’s ‘fitness’ is only part of the picture, however. As Haddadin points out, the emergence of AI has meant that robots can now develop and evolve, meaning that their capabilities can change. As an evolving resource itself, the Tree of Robots could prove essential in tracking these changes and teaching us how AI can optimize robotic machinery with new capabilities. 

“Take again the human analogy,” says Haddadin. “When we are born we can see the body, but we cannot yet tell what the capabilities are or might be in the future. That depends on learning — the human version of programming. And that’s a dynamic process; we learn throughout our lives, so our skills and capabilities change.  

“With the dawn of AI, this is also happening with machines. They are not just programmed once and this is what they do forever — no, now they evolve. The same embodiment might have fundamentally different capabilities based on the AI program running it. This new dynamic effectively creates new robot species and even genera merely based on software.”  

The Tree of Robots can play a crucial role in continuously reclassifying robots based on their new and emerging capabilities, rather than limiting our understanding of them to outdated assumptions about their design and function. And in doing so, it can help guide new research and development, accelerating the evolution of robotics and optimizing their potential.  

Creating a living encyclopaedia 

For the Tree of Robots to grow and flourish in such a way, there needs to be a concerted global effort with additions and contributions from around the world. And for this to happen, it was important to Haddadin and his fellow researchers that the framework was entirely made available open-source — ensuring that systems are evaluated with objectivity, reproducibility, and measurable benchmarks; creating what Haddadin believes will be the “living encyclopaedia” of the robotics field. 

“It sounds like a very grand vision, but I think this is where we’re heading to,” he says. “And for this to happen you need a global movement. Therefore, the concepts and data are shared, the protocols are shared, everything is very open-source — we don’t hide anything.  

“We’ve shown how the Tree of Robots works for basic manipulators, and it obviously now has to extend to all the other systems that are out there — humanoids, flying drones, space robots, underwater — you name it. To do that, it has to be as open as it can get so that it becomes as objective and encyclopaedic as possible. That way we get a transparent resource that is grounded and deeply rooted in measuring abilities and capabilities of systems so that they become comparable. 

“It’s a big effort. Robotics and AI have only recently become a true scientific field on their own. They have been very technology-driven, but in order to be a true science, reproducibility, openness of data, and objectification of results is at the core. We often say there’s a crisis of reproducibility in science, and creating efforts like this are, I think, helping to remove that crisis. Which in our field is as current as it gets.” 

These scientific processes, combined with its open-source framework, and in-depth classification system, certainly provide fertile ground for the Tree of Robots to thrive and bring ever-increasing clarity to the complexity of robots, their inner-workings, and their vast array of capabilities. As such, it will surely be a foundational resource for robotic understanding now and in the years ahead – and likely become a crucible for robotic and AI evolution, development and optimization.