01 You built a benchmark that tests whether AI models can safely control a humanoid robot near people. Why that topic, and why now?

“A robot is potentially a dangerous machine. We get a lot of robots here from different producers, including some from China, and now they all want to include some kind of agentic AI. In the news, we see people using ChatGPT, Claude or other models to operate a robot, and they do it without any safety concerns. So we asked ourselves: how will we handle these issues once these systems enter the market?

We are a research institute, so we want to give people a path to do this properly. The benchmark is on GitHub, and anyone can download it and test their own agentic AI system. The industry is already responding with new technology for running agentic systems, and we are keen to test it ourselves. I think the topic is even more relevant today than when I published it.”

02 For anyone who hasn't read the paper, what are the key findings?

“Whenever you give a robot a safety compliance list, you usually end up with over-compliance. That makes your robot safe, but not productive.

Here is a practical example. We told the robot that it should not move faster than 30 percent of its maximum speed if it sees a human. The smartest models worked out that they could keep working at 30 percent. The less capable models stopped doing anything at all. They were over-compliant with our original instruction.”

03 Is there a winner among the models you have tested so far?

“Currently, the more expensive one wins. But large language models that operate a robot have one problem: latency. I come from human–robot collaboration research. Even models that respond quite well are slow. If we want to interact with these robots, we have to use local LLMs, and those are less smart.”

04 Have you tested open-source models that run locally, such as Chinese models? How do they compare?

“Yes, and we get quite different results. In many cases, they cannot get past the hurdle of over-compliance. But these models are getting better and better.

Local LLMs have one nice feature: they respond, so they interact with you. If a robot tells me what it is going to do, that gives me feedback. It is not a traditional safety element built into a safety layer. But it is how we as humans communicate.”

05 What do these findings mean for companies that want to put AI-driven robots around people?

“The crux is whether they are allowed to use these models in an environment where real people are present. An LLM can give a naughty answer, and that won't harm you. But a robot is potentially a dangerous machine.

There are a lot of legal implications. You cannot just run a robot with some kind of LLM in an environment where people are, unless in the end you take ownership of that operation and are legally accountable for it.”

06 What perspective will you bring to the panel in Munich?

“I am very passionate about humanoid robotics. I truly believe it will be one of the biggest industries of the next ten years, and possibly the next fifty.

My work is not only LLM testing. We also run physical tests of real applications, and so far we have tested five different humanoid robots in our facility. At Fraunhofer IPA we are building up a lab where we can not only test these robots but also operate them and bring them into the real market. The goal is that SMEs in German industry can use these robots to their advantage, so we work closely with industry.”

07 The summit brings European and Chinese OEMs, integrators and buyers into one room. Who are you most keen to meet?

“I want to meet them all. There is a big race going on. So far I have mostly worked with robots from Dobot, Unitree and AgiBot, and from Germany, with Neura. Everybody is really pushing, and I expect some of them to pull ahead. But I want to get my hands on all of these robots, so I will probably meet them all.”

08 Finally: a student wants to get into Physical AI and robotics. Where should they start?

“The open-source community already gives you everything you need. Download the software from GitHub and test it in a simulator such as Isaac Sim. Use the webcam in your laptop for some simple computer vision, because that is how your robot will see. Once you can see a box, at some point you can pick it up. Start in simulation, and then move to a real robot.”

Meet Aulon Bajrami in Munich

He speaks on a panel at Humanoid Europe 2026 – Europe & China Humanoid Robot and Physical AI Summit, 5–6 November 2026, Munich. Register now →

The interview has been edited for length and clarity. The views expressed are the interviewee's own.