Building Industrial AI for the shop floor

The Future-Ready Podcast: Industry & Beyond ·

Show notes

Robotics have been a staple in manufacturing for decades, allowing certain repetitive tasks to be automated yet, for as powerful as automation is, replicating full human-like abilities is difficult to achieve. By incorporating AI with advanced robotics, a new level of flexible, adaptable, robotic systems can not only take on tasks previously limited to humans, but also support rapid product iteration and development in a way never before possible.

In this episode, host Spencer Acain is joined by Christopher Schuette, Senior Product Portfolio Manager for Robotics AI at Siemens Digital Industries to examine what it takes to build production ready Industrial AI solutions capable of operating complex machines under a variety of conditions.

In the episode you’ll learn about:

(2:12) Upgrading to AI-powered robotics

(4:23) Building more dynamic robotics with AI

(9:41) Training AI for the shop floor

(12:58) Data needs for production ready AI

About the voices:

Christopher Schuette is the Senior Product Portfolio Manager for Robotics AI at Siemens Digital Industries

Spencer Acain is a technical marketing writer for AI thought leadership at Siemens Digital Industries with a background in Applied Mathematics and Mechanical Engineering.

Transcript

Hello and welcome to the AI Spectrum podcast. I'm your host, Spencer Kane. In this series, we explore a wide range of AI topics from all across Siemens and how they apply to different technologies. Today I'm joined by Christopher Shutte, senior Product portfolio Manager for robotics AI at Siemens Digital-Industries. Welcome, Christopher. Thanks, Spencer. Great to be here and hello to all listeners. To start things off, uh Christopher, could you tell me a bit more about yourself and your work at Siemens? Yeah, for sure. Um, so I'm working for Siemens uh quite a while, um, starting directly after my study of megatronics. Um, I started in customer services. I wanted to explore, uh, the fields out there to how users are. What they're doing with our products, the Siemens products, um, and after that, also some time in system testing, product testing, and also training our customers in new products, um, I turned to business development to just follow the question. How's the future of automation look like, um, and can I shape it or be part of it and, uh, one of these parts in business development was great to be at the birth of a new vertical that is called vertical in logistics, uh, what is an amazing field of, uh, new technologies, the shuttle systems, HEVs, robotics, um, all these great things that happened there and that brought me to an, uh, own initiated. Project, uh, that turned now to a complete product line, um, where we started in to exploring the field of AI enabled robotics, um, and as you said, uh, I'm now the product portfolio manager in this field of robotics, AI products. It focused on AI powered applications that you can upgrade off the shelf robots with. Great. So it sounds like you've really been in this from the start. Um, can you tell us a bit more about those, you know, AI powered robotics, you know, how are you upgrading robotics? What does that look like? And um you know, why is this technology important? Oh yeah, um. Why it's important, so we identified that, uh, I take some studies so make also the word for tell us that around 50 to 60% of work in manufacturing are still performed by manual labor forces and are not fully roboticized or automated and around 50 to 40% are just in dynamic changing environments and where traditional automation cannot catch up it's widget. So it struggles in performance and for that reason humans still need to perform repetitive robot-like work in a time where the labor pool of workers are shrinking, especially in industrial countries, and the payment of that workers are rising, so the economical pressure is high. That are the foundation of what we're doing, and what we're doing is um. So our first product is probably the one of the first in the area of physical AI, um, that terms try to explain or differentiate the digital AI that is behind your screen, um, that helps you to navigate, that helps you to search in search algorithm. Uh, and also, uh, the chat GPT is for sure something like in this area, but the physical AI touches our real world, so it's intact with the real world, um, and it sees reason and act in dynamic changing environments. That is a complete new paradigm if this realized as we think it will be. And the first product that we bring up was in the portfolio, uh PAI that enables any robot to pick items that it has never seen before. Uh, without training by the user, very easy to use, very easy to integrate in existing machines, new machines, and that was just our start point. Yeah. So it sounds really like you're, you're kind of addressing that core issue like you're saying of, you know, robots are very static in the past and the way they can pick stuff. So you have this, this, um, You know, just ability to handle whatever at any time. And then, um, but how is that from your ability to leverage like more advanced AI now, like, how are you seeing the technologies, that type of technology improving with, you know, with more powerful AI accelerators available because surely this is something that's um that requires that kind of extra power to be able to just do that humanlike tasks without without much training, without any training. Uh, you know, like, for example, you know, I know Siemens start to integrate like Nvidia accelerators in their PLCs. Is this something you're able to take advantage of? Oh yeah, definitely, Spencer. Um, first of all, we learned very early that we in industry want, yeah, we have to sync in industry in system values, not in isolated technologies. To your question, so, um, combined PLC, um, and it's fact that the most of the industrial process out there are today PLC based, um, needs to somehow merge or integrate with what is possible for AI today that is powered by GPUs also for an Nvidia. And um for that, it's very good to see that Siemens and Vidia joining forces, building rating systems, and that it merging closer and closer together. So we, we make some experience to have GPU power technology very close to the PLC. Um, and then we learned also to have it separate in the PC is also worth it, and then have a better interfaces. But that is the moment of progress, and I'm very happy that we are also driving the way toward a software defined automation concept. And the idea behind that, it's to bring Um, automation to a level where it comes from a one-off specialized machine for one customer to amoized product that you can scale out on the market with less engineering effort, and you can achieve that. By having the ITOT congregation, um, you take the engineering methods from IT that was built over decades and converted to the industrial level. That is what we're doing with products like um. The automation expansion product that is giving you an IT like engineering. On the other hand, to answer your question on the GPU accelerator level, uh, we have more and more products here that taking that type of hardware combining better to the PLC world and also to the level where the PLC get visualized, um. And you can put that on one device, but also great, and also you can deploy that with a concept like industrial edge. So all these things are perfect and it's really a system value, but I want to tell you what we learned. To use bigger AI models. So our R&D teams come from, Um, AI based computer vision for post estimation at the very beginning. And we released just last year a second product tier um that use also an industrial grade foundation model for segmentation of random items as is one part of our product and it helps to provide our users with more information about the unknown item. So at least the length, the width, the orientation, it also part of an intelligent suction group activation feature that adapts to the multi-suction group are used. To adapt to the item size position in the room and at the end it's for the user less to change ah and picking in a very wide range of items just by advancing the architecture of neural networks. And we have another core metric that is the end to end compute time. And we learned with the huge foundation model that I just named. It took us from some milliseconds, so the first A tier of our PI was able to run on a, yeah, tablet PC like CPU without the GPU, that was OK. The new version edition needs a GPU because it takes with GPU over 2 seconds to compute. That is not good, uh, when you have a very high performance machine. So you need this GPU integration to reduce the performance and to compute time heavily. That is obvious. I'm very great to have us so integrated with Nvidia together, um, and it's so great that we are getting more and more friends here with Nvidia colleagues. We are in regular exchange and, uh, bringing this infrastructure in a more tangible system value way on the market. Um, very curious about what will happen also in the future here. Yeah, no, I can, I can imagine, but kind of to. To pivot from that, like, you know, you, you have all this extra power available now in like these PM the PLCs on the factory floor with these accelerators, but to reach the kind of goal that you're talking about of just being able to have a robot that's picking anything and everything without any training at the end or very little training at the end, is that, um, how much like back end training has to go into these AI models? Is it all? Like, do you pre-train it completely and then just ship it out and then the customers don't need to do anything else, or do they have to train as well? Or how are you able to, how are you managing that and that kind of training of these models to be able to handle these diverse situations? Unfortunately, I can give not a very clear answer to that, but let me try it. Um, one of the fundamental product decisions, uh, we did, it's, if you want to give these retraining capability in the hands of our users, uh, which are integrators and OEMs of, so these creators of advanced robotics solutions, um, with all its complexity, um, so they need to to get the data, they need to label it, retrain it, test it. Um, or if we make their lives easier, as easy as possible and keep the retraining in our product team. And we decided to take it in our product team and give the OEMs integrators an out of the box performance model and in the case the model not reaches the performance that we expect and theirs expect, uh, we are able to retrain these models. And how big is this effort to do so, just to give you a thought. We haven't Product that uh the the value proposition is that it can grasp uh the most of the items in the warehouse nearly to every item in the warehouse, um, we are on the way to, to get there. It's different to these approaches where you specifically train one item in manufacturing probably in a new network and it, it works reliable for that one out for another one. It's especially why we take these training in our capability in a product team because we cover item classes, not a specific product that is fitted by CAT data, for example. And how big is such an effort? It took us up to 3 years. To come from a little baby network um for 5 degrees of reasons, uh pick post detection and computation with a really good industrial grade performance to what we have today, um, a multi-section we ready, uh, deployed in the field, uh, neural network, the foundational model backed, um, that really makes a job in, in operations. So 3 years of collecting data, firstly from, uh, synthetic data from simulation, because it's cheap to generate and easy to generate, but it turns out that you need to find other ways to make really a reliable industrial grid. So, beyond that synthetic data you were just talking about, you know, how are you able to get all the training data then, um, you know, obviously, like you said, it can't, you can't just use CAD data or synthetic data for this, uh, so, so you needed a lot of real data. To pre-train and develop your models, where, so where are you able to get all of that? Yeah, good question. The underlying challenge most companies working on physical AI are facing is the limited availability of large scale real world robotics data. Models like GPT were trained on internet scale data sets, and we don't have this equivalent data foundation for robotic learning. So the data gap remains very large and it's probably one of the biggest challenges for model performance across the market and its success. So it's well recognized across the technology field. For applied AI robotic system in industrial settings, continuous improvement depends heavily on operational real world data, particularly on the edge cases. Uh, this includes vision data along with contextual signals such as system states, also from the PLC and outcomes. For example, perception foundation models like segmentation models, foundational ones are trained on large graded labeled data sets, often still involving significant manual annotation efforts. Yeah, looking ahead, vision action models promise to generalize manipulation skills across tasks and environments and to achieve this, virtual training data sets are required. Today, a combination of visual observation with trajectory, level, motion, context, um, this allows models to understand the scene versus grade and adapt directly the robot motion to it. Um, the technology field is evolving quickly. Within a year, entire foundation model architectures can change several times, um, driven also by new research insights and this significantly advanced the maturity of the technology. Um, one constant will probably stay, uh, infrastructure to feed the need for large scale data collection and continuous moral improvement. And it's this vision behind the mostly automated data flywheel concept. In the foreseeable future, real world data stay a must, not optional, and to achieve internet great performance. That is why we actively building the capability within our product team. Yeah, I mean, obviously, building that kind of infrastructure and designing that isn't something you want to have everybody doing on their own all the time, but that is all the time we have for this episode. So once again, I have been your host, Spencer Kane, joined by Christopher Shita on the AI Spectrum podcast. Tune in again next time as we continue diving into what it takes to bring AI automation into the factory.

Full transcript (WebVTT)