Bench Talk

Rafik Mitry from Mouser Electronics Interviews Professor Alois Knoll

(Source: Andrey Suslov/Shutterstock.com)

In September, I chatted with Professor Alois Knoll, Technical University of Munich’s chair of the robotics and embedded systems department, about artificial intelligence (AI). Knoll’s research interests focus on human-robot interaction, service robotics, medical robotics, cognitive robotics, and cyber-physical/embedded systems.

In our conversation that follows, Knoll covered his current research, the challenges of Edge AI, Neuralink, and automotive applications.

What is your research focus in robotics currently?

Currently, we are focusing on:

• Making AI more secure and reliable
• Using AI in critical infrastructure
• Optimizing production processes through AI

We are also working on two main research projects: the Human Brain Project (HBP) and Roboy.

Human Brain Project

The Human Brain Project is a very big project where brain research and technology development come together. Two areas that receive a lot of public attention in the Human Brain Project are neuromorphic processors and the other is neurorobotics. The Human Brain Project builds research infrastructure to advance neuroscience, medicine, and computing. It is one of the largest scientific projects funded by the European Union and one of the four Future and Emerging Technology Flagships. The project began in 2013 and will go on for 10 years. It has around 800 scientists at more than 100 universities, teaching hospitals, and research entities across Europe.

The project's main goal is to study the multi-level complex biology of the human brain: how it's built and how the neurons function. This acquired knowledge should then be deployed in brain-derived applications in health, computing, and technology.

Roboy Project

Roboy is a project that receives a lot of public attention because it is one of the most advanced robots in the world in terms of its objective, which is the embodiment of an AI. It's an advanced humanoid robot that is modeled on the musculoskeletal system of the human body. Roboy has muscles and tendons that differentiate it from the traditional robots that have motors in their joints. The Roboy team at the TUM is developing their cognitive system with a dialog system, speech-to-text, text-to-speech, and a memory system. The first version of Roboy was built in March 2013, and by 2019, Roboy was able to sell ice cream. The plan is by 2050 that Roboy will be as good as a human.

What kind of challenges are we facing with Edge AI right now? Are they on the software side or the hardware side (compute engine) or ethical (bias)?

The ethical aspect is one that we always have to be aware of. That is why we have a group in the Human Brain Project that only deals with ethical questions and the ethical aspect is taken into consideration. The basics are that the decoding of the human brain is still in the early stages. The implementation of what we know is easy, that's because the technology is already very advanced. This leads to chips that are already commercially available or even available on a laboratory scale. Still, we can expect that neuromorphic processors will play an essential role in the application of AI in the future. For example, there is the SpiNNaker system development, which is based on numerical models running in real-time on custom digital multicore chips using the Arm architecture. This system performs aspects of biological neural networks as analog or digital copies on electronic circuits. The SpiNNaker system features 30,000 custom digital chips, each with 18 cores, resulting in a total of over 500,000 cores. It also has a shared local 128Mbyte RAM. Then there are commercial neuromorphic processors, e.g., Intel's Loihi. The speed on the development side of the hardware is really impressive.

Do you think that something could be achieved with Neuralink?

(Tesla CEO) Elon Musk presented a biocompatible chip that gets connected to the human brain's neurons, which is significant progress in medical research. If its purpose is to cure certain illnesses, e.g., with Parkinson's, Neuralink is, in fact, promising. But, if your expectation is that you can play music into the human brain or make humans more intelligent, I believe it is unrealistic. Nevertheless, this is a very exciting development from a medical perspective for helping the market cure brain diseases.

Why can’t AI in automotive applications make self-decisions on new events?

There are different approaches to bringing AI into the car. One method is the so-called end-to-end approach, driven by Nvidia, where you simply let the car drive in different scenarios. This can be done, but has some difficulties:

• Generalization ability: If another scenario comes up that this car has not seen yet, then it does not know what to do.
• The second difficulty is that of traceability. When the AI makes a decision, it is not possible to understand why it reacted in that way and made this decision.

For example, while on a highway, AI can easily do most of the tasks when the proper parameters are up and running. However, when you leave the highway and enter the city, the traffic becomes very complex.

If a problem shows up that the AI does not know—because it did not appear during model training—this is still a major problem for which there is no solution at the moment. Therefore, there are no fully autonomous cars on the market until today. This is due to the methods currently being used. We still do not know how to develop a neural network architecture that will enable fully autonomous vehicles. This is not primarily software; it is the question of methodology and structure.

Finally, what advice can you give to students who want to work in AI?

Come to the Technical University of Munich Master Program Robotics, Cognition and Artificial Intelligence. You will get an in-depth overview of the field and an excellent entry into the industrial market, both in robotics and in the AI side.