It is probably happening right now

In the course of the presentation of the Universal Robot UR20e, which differs from its older siblings by a completely different inner workings, the supplier Synapticon (Stuttgart) was presented here, among others. Without such innovative suppliers, some innovations would not even be possible, according to Summer 2022 CEO Nikolai Ensslen, who has outstanding knowledge of the market and technology. His numerous business trips around the globe have certainly contributed to this. Without a doubt, Nikolai Ensslen can be described as a thought leader in the industry. I am pleased to welcome him as a guest author here. The post fits well with yesterday's about OpenAI's entry into 1X Raises.

In recent years, I have often been asked for my opinion on the state of development of artificial intelligence, usually in the context of robotics. My usual response has been that while AI in robotics has helped bring us to a solid level of autonomous navigation and motion planning, object recognition and manipulation, and human-robot interaction (speech), truly intelligent robot behavior, reasoning, and decision making are still a long way off. While all of the aforementioned "infrastructural AI functions" have been working well for more than a decade, their processing speed evolved from "frighteningly slow" to "acceptable" during that time.

Inspired by science fiction for 80 years, impressed by primarily Japanese research more than 30 years ago, and stunned by viral videos of seemingly very capable robots over the past decade, people are wondering when robots will finally become what we've long imagined them to be. After the revelations of the past two weeks, I've come to the conclusion that there's a good chance we'll be there soon - if not very soon.

OpenAI released GPT-4, announced plug-in support for ChatGPT, and 1X (formerly Halodi) announced that it is funded by OpenAI and others - which seems like a nod to me: Few have yet recognized the potential impact of large-scale language models on physical robotics.

When ChatGPT was released late last year, I could envision how useful LLM technology could be in creating better human-robot interfaces, such as NEURA Robotics had introduced almost two years earlier with their MAiRA series. Although I was working on glueware and HW accelerators for AI technology components when I entered the robotics industry about 14 years ago, I didn't realize how far beyond that the role of LLMs in robotics could go. The GPT-4's image input finally tipped the scales in my favor.

If GPT is able to explain why some nerdy gadget is funny, it is probably also able to understand that, for example, an unconscious person needs help in an accident. We'll see how far this can go, and there's still some work to be done to meaningfully return the generated language as input to existing robotic software stacks. Still, I'm hopeful here: ChatGPT based on 3.5 already wasn't bad at writing ROS nodes. What is certain is that it can make robots MUCH more intelligent than they are today.

In contrast to the implications for robotics, there has been some public debate about whether ChatGPT can be considered AGI (artificial general intelligence, i.e., human-like intelligence). My opinion on this is: No, not yet. But ChatGPT has made it very clear what a smooth and difficult-to-assess transition it will be from purpose-built AI/components to AGI. And the insights of the last few weeks have sharpened my judgment about the importance of AGI for robotics.

When asked about the emerging hype around humanoid robots in recent months, I shared my prediction that humanoid robots need AGI for their use and adoption to take off - more rationally, for them to make sense as commercial products. Purpose-built robots will prevail as long as the software is also purpose-built. Once the software becomes universal, so must the robot, according to my simple logic.

There will be more high-performance drives in a humanoid robot than in a car

The high cost of humanoid hardware will not come down to the desired level, the price of an affordable car, until the industry also experiences adoption on the scale of cars. The electrotechnical complexity of a humanoid can be considered even higher than that of a car: many more high-power drives are required in a humanoid than in a car, and the remaining device complexity is similar, although naturally lower in volume.

I now believe that GPT-4 (and similar technologies) will enable robotic software to be so universal that universal physical robotic bodies will also be useful. Therefore, I now strongly believe that the humanoid wave we are currently experiencing will be sustainable. The available software technology will soon make it possible to develop "sufficiently useful" robotic products to create a market for at least early adopters.

At the same time, we need to keep in mind that this is where the extreme pace of generative AI meets the usual harsh reality of physical robotics - which is, unfortunately, much slower than the rapid succession of news we have become accustomed to in "IT AI". Mechanical, electromechanical, mechatronic and embedded technology is simply much harder to master than things from the IT world. That's why there are fewer people with the necessary expertise. And because of the higher risk, there's a lot less money flying around than there is. Maybe, hopefully, this is another event to change that.

On the other hand: When the virtual world meets the physical world, it will probably be much more difficult, and for good reason. Some are already fearful of recent advances in AI (and rightly so). The recent introduction of plug-ins for ChatGPT is seen by experts as a big risk-taking move by OpenAI - and a breach of their promise to be a cautious AI player. How much should we fear when these plug-ins go mechatronic?

Quite a bit. The way to get there: a new industry, comparable to the certification-intensive automotive, aerospace or medical device industries, is emerging. Fortunately, with the advent of collaborative robotics and autonomous mobile robots in recent years, certified functional safety has become an increasingly central issue in industrial robotics, and the industry is just getting to the point where full certification of safe motion monitoring and limiting is becoming a must.

As the head of the company with the world's only fully safety-certified motion control product portfolio for robotics, I know what is possible today when it comes to making the robots we have today as commercial products safe. Certifiable functional safety must be thought of much more broadly than certifiable behavioral safety once universal humanoid robots are to be released to the public.

With every step AI takes from now on, robotics will have to make three efforts. Let's roll up our sleeves, exciting times are coming!

In the video below, customer Robco explains its decision to use Synapticon.