# Exploring Physical AI: Merck’s Journey from Humanoid Proof of Concept to Industrial Value (Presented by EY)

## Abstract

Key takeaways Learn how digital twins and simulation-first development shorten deployment timelines and reduce implementation risk for physical AI applications. Understand why scaling robotics requires more than robots—success depends on reusable skills, governance, safety frameworks, and enterprise integration. Explore how hardware-agnostic robot orchestration enables long-term flexibility while protecting investments in skills and software. Discover how Merck is building a roadmap from a pick-and-pack use case to a scalable physical AI platform that amplifies—not replaces—our workforce worldwide. Apply practical lessons on safety-by-design, digital validation, and organizational readiness when deploying physical AI in regulated industrial environments.

## AI Summary

_No AI summary available._

## Transcript

**[00:00:05 – 00:00:11]** All right, good morning everyone, and welcome to our session.

**[00:00:11 – 00:00:16]** Today I would like to bring you and take you on a journey,

**[00:00:17 – 00:00:23]** on our own journey into physical AI, and at the same time would

**[00:00:23 – 00:00:26]** like to share with you not only what experience and what

**[00:00:26 – 00:00:30]** we're doing, but particularly also how we assess the value,

**[00:00:30 – 00:00:34]** the impact, and the potential.

**[00:00:37 – 00:00:43]** To support the incredible Merck team, we partnered with EY as an

**[00:00:43 – 00:00:48]** integrator and implementer, leading it, and at the same time with

**[00:00:48 – 00:00:55]** our trusted technology partners, Bundlebots and Grid Dynamics.

**[00:00:55 – 00:00:59]** Together, we have achieved so much only within a couple

**[00:00:59 – 00:01:03]** of months, and I'm incredibly proud to be here, standing here today, to

**[00:01:03 – 00:01:07]** present what the team has achieved.

**[00:01:07 – 00:01:10]** So, before we get started, I would like to share with

**[00:01:10 – 00:01:13]** you who Merck is.

**[00:01:13 – 00:01:16]** So, as you may know, there are two different companies with the same

**[00:01:16 – 00:01:20]** name, so let me clarify who we are.

**[00:01:20 – 00:01:25]** We, Merck, are known as Merck internationally, but also known as

**[00:01:25 – 00:01:33]** EMD, Serono, Millipore Sigma, and EMD Electronics in US and Canada.

**[00:01:33 – 00:01:37]** We are 360 years old, so you can imagine we have been going

**[00:01:37 – 00:01:40]** through a couple of different industrial revolutions over

**[00:01:40 – 00:01:44]** the last years, and we managed to not only sense the right

**[00:01:44 – 00:01:48]** signals at the right moment, but also to identify the right partners

**[00:01:48 – 00:01:54]** who help us to transform ourselves and to address our business needs.

**[00:01:55 – 00:01:56]** In total, we are talking about 320,000 products that we deliver

**[00:01:56 – 00:02:00]** to 1.5 million customers and 100 million patients.

**[00:02:04 – 00:02:06]** We have three different business sectors.

**[00:02:06 – 00:02:09]** The oldest one, we are the oldest pharmaceutical company

**[00:02:09 – 00:02:12]** in the world, for those who don't know, we call it healthcare sector.

**[00:02:12 – 00:02:16]** The second, Millipore Sigma, with LifeScience, which is

**[00:02:16 – 00:02:19]** having 300,000 products for the LifeScience industry,

**[00:02:20 – 00:02:23]** and Electronics, our service and raw materials provider

**[00:02:23 – 00:02:25]** for the semiconductor industry.

**[00:02:25 – 00:02:28]** So, having said that, you can imagine operations is

**[00:02:28 – 00:02:30]** quite a big deal for us.

**[00:02:30 – 00:02:33]** It's quite important to deliver the products to our customers

**[00:02:33 – 00:02:39]** in the best quality, in time, and, of course, in an affordable manner.

**[00:02:40 – 00:02:42]** So, what is the challenge that we're facing at the moment?

**[00:02:42 – 00:02:46]** Let's talk about challenges that we're facing as an industry.

**[00:02:46 – 00:02:50]** So, the demographic change is a reality.

**[00:02:50 – 00:02:54]** It's a reality that's actually coming to our shop floor.

**[00:02:54 – 00:02:57]** It's a reality that we are all facing as an industry.

**[00:02:57 – 00:03:04]** Indeed, we see that in Europe, the working age workforce is reducing

**[00:03:04 – 00:03:07]** over the next couple of years.

**[00:03:07 – 00:03:10]** This is one of the elements that we have identified already

**[00:03:10 – 00:03:15]** as a big key challenge, and it's already affecting us.

**[00:03:15 – 00:03:18]** At the same time, a lot of this manufacturing process

**[00:03:18 – 00:03:21]** steps are still manual.

**[00:03:21 – 00:03:26]** So, we see the gap that will exist, as the work will continue to exist.

**[00:03:26 – 00:03:30]** The second big element that we see is that the automation so

**[00:03:30 – 00:03:36]** far has really helped us a lot in a space where we have high volume but

**[00:03:36 – 00:03:39]** less of complexity in the product.

**[00:03:39 – 00:03:43]** Now, if we're adding now into a space where actually reducing

**[00:03:43 – 00:03:46]** the volume and increasing the flexibility of the product

**[00:03:46 – 00:03:49]** and diversity of the product, well, that is the moment where

**[00:03:49 – 00:03:52]** the classical automation may not help us anymore.

**[00:03:52 – 00:03:55]** We need to enter a space where the agility of robots and

**[00:03:55 – 00:03:58]** automation needs to enter, and that's actually now the

**[00:03:58 – 00:04:01]** space that we also identified as a new trend, where the

**[00:04:01 – 00:04:08]** convergence of robots, automation, and AI is coming nicely together.

**[00:04:08 – 00:04:09]** It's maturing.

**[00:04:09 – 00:04:13]** Not only the AI models, but also the respective skills

**[00:04:13 – 00:04:17]** is coming to a maturity where the convergence of this with the

**[00:04:17 – 00:04:21]** business need that we identify at Merck is actually providing us the

**[00:04:21 – 00:04:27]** right moment to focus on humanoid robots and how they can really help

**[00:04:27 – 00:04:31]** us on our challenges in operations.

**[00:04:31 – 00:04:35]** With that, I would like to hand over to you, Daniel, to go next.

**[00:04:35 – 00:04:37]** Thank you, Michelangelo.

**[00:04:37 – 00:04:42]** So, the question now, as we're coming to the heart of the

**[00:04:42 – 00:04:47]** project, was, how can we transfer exactly this ambitious vision

**[00:04:47 – 00:04:50]** into a use case, which is industrialized on the one

**[00:04:50 – 00:04:55]** hand, shows the entire capabilities or potential of the project

**[00:04:55 – 00:05:00]** and the robot, and which has at the same time a high potential

**[00:05:00 – 00:05:04]** for the most and significant scalability, industrialized?

**[00:05:04 – 00:05:09]** So, we decided for a pick-and-pick process.

**[00:05:09 – 00:05:13]** Say, okay, pick-and-pick process, very simple now, but if you

**[00:05:13 – 00:05:18]** break it down into seven skills, for example, you see the different

**[00:05:18 – 00:05:22]** skills and capabilities a robot needs at this stage, starting

**[00:05:22 – 00:05:25]** with folding a cartoon box.

**[00:05:25 – 00:05:29]** Very high sophisticated because we are far from linear physics

**[00:05:29 – 00:05:33]** and we need a lot of experience, still state of the art.

**[00:05:33 – 00:05:40]** We need capabilities in picking up things, getting a vision,

**[00:05:40 – 00:05:44]** proof of the things are right, the vision set aside.

**[00:05:44 – 00:05:49]** We also need to pick it back, have certain skills, and if

**[00:05:49 – 00:05:53]** you break it down into the seven, you see all high sophisticated

**[00:05:53 – 00:05:59]** topics, besides also some low sophisticated topics, are covered.

**[00:05:59 – 00:06:01]** Why is this so important to us?

**[00:06:01 – 00:06:04]** We didn't want to be just another player playing

**[00:06:04 – 00:06:05]** around with robotics.

**[00:06:05 – 00:06:09]** We want to make it meaningful and reliable that we get it

**[00:06:09 – 00:06:11]** from the first day, where we say this is the way we

**[00:06:12 – 00:06:16]** want to go, to a scalable case.

**[00:06:17 – 00:06:23]** Saying this, robotics turned very quickly out to be a team sport.

**[00:06:23 – 00:06:28]** We see NVIDIA, we see bundle bots, we see grid dynamics, but none of

**[00:06:28 – 00:06:35]** them by their own can realize the robot and the task he is executing.

**[00:06:36 – 00:06:38]** We have Wonderbots who are presenting an independent

**[00:06:38 – 00:06:43]** platform because for us it was very important, as the hardware

**[00:06:43 – 00:06:48]** market is very high developing, let's say in this way, that we

**[00:06:48 – 00:06:51]** also can reuse the trained skills.

**[00:06:51 – 00:06:54]** For Grid Dynamics, we use the entire tools and also

**[00:06:54 – 00:06:59]** the deep experience in AI and physical AI and robotics.

**[00:06:59 – 00:07:02]** And last but not least, NVIDIA, an absolutely game changer

**[00:07:02 – 00:07:05]** at this point, because now everything what makes sense comes

**[00:07:05 – 00:07:08]** together, also with the Omniverse, with the entire tool suite,

**[00:07:09 – 00:07:11]** to realize what we want to do.

**[00:07:11 – 00:07:16]** And I just can say it's really a team sport to bring the

**[00:07:16 – 00:07:18]** best in class together.

**[00:07:18 – 00:07:23]** And I think today, in 2026, we have now the maturity and

**[00:07:23 – 00:07:28]** the experience to bring such ideas and visions into life.

**[00:07:28 – 00:07:31]** Saying that, handing over to the team who performed

**[00:07:31 – 00:07:34]** the task, Jan Niklas.

**[00:07:34 – 00:07:36]** Thank you, Daniel.

**[00:07:36 – 00:07:40]** One fundamental principle of our approach is to invest

**[00:07:40 – 00:07:44]** in robotic capabilities and not in individual robot generations, and

**[00:07:44 – 00:07:45]** the reason for that is very simple.

**[00:07:46 – 00:07:48]** The humanoid robot market is emerging

**[00:07:48 – 00:07:50]** incredibly fast these days.

**[00:07:50 – 00:07:54]** So, if we introduce a new robot to a factory, we want to ensure that

**[00:07:54 – 00:07:59]** we do not start from scratch every time we introduce a new robot.

**[00:07:59 – 00:08:03]** Instead, we want to build those skills that we develop as reusable

**[00:08:03 – 00:08:07]** and as transferable as possible between the robots, and for that,

**[00:08:07 – 00:08:10]** BundleBots Nova comes into play.

**[00:08:10 – 00:08:12]** We distinguish between three different layers.

**[00:08:12 – 00:08:15]** On top, you have the reusable skill library.

**[00:08:15 – 00:08:19]** Those are the capabilities that we develop along the project,

**[00:08:19 – 00:08:23]** that we continuously improve, and that we want to make transferable

**[00:08:23 – 00:08:25]** between the different robots.

**[00:08:26 – 00:08:29]** Those can be capabilities such as object detection,

**[00:08:29 – 00:08:34]** navigation, distribution, inspection, and bringing that

**[00:08:34 – 00:08:38]** to life, this is where Wonderbots Nova comes into play.

**[00:08:38 – 00:08:41]** So, Nova is our hardware agnostic orchestration layer, which

**[00:08:41 – 00:08:45]** brings capabilities such as motion control, skill execution, but

**[00:08:46 – 00:08:50]** also the communication to different underlying hardware systems.

**[00:08:50 – 00:08:54]** So, NOVA builds our connection system between the skill library

**[00:08:54 – 00:08:55]** and the different robots.

**[00:08:56 – 00:08:58]** And on the bottom, we have those robots.

**[00:08:58 – 00:09:02]** Those can be humanoid robots, as in our case, AGI-Bot G2

**[00:09:02 – 00:09:06]** and AGI-Bot A3-Ultra, that we're benchmarking on this project.

**[00:09:06 – 00:09:10]** But those can be also industrial robots, and those can be also

**[00:09:10 – 00:09:14]** mobile robotic solutions, AGVs, robot dogs, for example.

**[00:09:14 – 00:09:18]** Of course, hardware independence does not mean that we can

**[00:09:18 – 00:09:22]** transfer those skills directly and each skill will work

**[00:09:22 – 00:09:24]** immediately on each robot.

**[00:09:24 – 00:09:27]** But the idea and the ambition is to transfer as much of the

**[00:09:27 – 00:09:33]** developed intelligence as possible and make those skills reusable.

**[00:09:34 – 00:09:36]** So, our principle is very simple.

**[00:09:36 – 00:09:40]** Skills stay, and robots can change.

**[00:09:42 – 00:09:44]** Now, once we have defined the skills that we want to

**[00:09:44 – 00:09:47]** develop, the next question is, how do we actually teach

**[00:09:47 – 00:09:48]** the robot to perform them?

**[00:09:48 – 00:09:51]** And what I can already say is, there is no single training

**[00:09:51 – 00:09:54]** approach that works equally well for every task.

**[00:09:54 – 00:09:58]** So, in our project, we're combining two different approaches.

**[00:09:58 – 00:10:02]** The first one is simulation first, where we combine conventional

**[00:10:02 – 00:10:06]** machine learning and reinforcement learning, and the simulation-first

**[00:10:06 – 00:10:12]** approach works best for tasks that have manageable physics to handle.

**[00:10:12 – 00:10:16]** So, for example, locomotion, palletizing, or just

**[00:10:16 – 00:10:18]** handling rigid objects.

**[00:10:18 – 00:10:22]** For that, we use NVIDIA technology, which is Omniverse, IsaacSim,

**[00:10:22 – 00:10:26]** IsaacLab, to create virtual environments, to train the

**[00:10:26 – 00:10:29]** skills, to develop the skills in the simulation, and this

**[00:10:29 – 00:10:33]** brings one major advantage, which is scalability.

**[00:10:33 – 00:10:37]** We can train across various data, we can generate large amounts

**[00:10:37 – 00:10:41]** of training data, we can introduce many variations and train different

**[00:10:41 – 00:10:45]** simulation runs in parallel, and all of that without exposing

**[00:10:45 – 00:10:48]** the hardware to unnecessary risks.

**[00:10:48 – 00:10:51]** Of course, simulation also has limitations, and this brings

**[00:10:51 – 00:10:54]** me to our second approach, which is the real-world first approach,

**[00:10:54 – 00:10:57]** where we use imitation learning.

**[00:10:57 – 00:11:02]** This is mainly meant for a task where we handle complex physics.

**[00:11:02 – 00:11:04]** Think about, for example, folding a carton box.

**[00:11:04 – 00:11:08]** Think about positioning a label or a sticky packaging tape on a box.

**[00:11:08 – 00:11:13]** Those are tasks with very complex materials, very complex physical

**[00:11:13 – 00:11:16]** interactions, and this, this is why we focus on tail operation

**[00:11:16 – 00:11:18]** directly on the physical robot.

**[00:11:18 – 00:11:22]** So, we collect this data, this demonstration data, use

**[00:11:22 – 00:11:27]** this resulting data to train the robot, and the robot can

**[00:11:27 – 00:11:32]** perform the task autonomously after dedicated runs, training runs.

**[00:11:32 – 00:11:36]** The disadvantage of this approach is, this is very time intensive,

**[00:11:36 – 00:11:38]** so we need to find the right mix between both approaches,

**[00:11:38 – 00:11:42]** and this is one great outcome of our project so far, that we need

**[00:11:42 – 00:11:48]** to mix and match according to the physical intensity of each task.

**[00:11:50 – 00:11:53]** Having worked on this project over the past month, I would

**[00:11:53 – 00:11:55]** like to share three lessons that we have learned so far,

**[00:11:55 – 00:11:59]** and I think those lessons are particularly relevant for every

**[00:11:59 – 00:12:01]** company who would like to start their own physical AI journey.

**[00:12:01 – 00:12:04]** So, let's start with lesson number one.

**[00:12:04 – 00:12:06]** Match the training approach to the physics of the tasks,

**[00:12:06 – 00:12:08]** and this I just discussed.

**[00:12:08 – 00:12:11]** So, simulation is very powerful if we handle tasks with

**[00:12:11 – 00:12:13]** manageable physics.

**[00:12:13 – 00:12:16]** But if it comes to complex materials, complex physics,

**[00:12:16 – 00:12:20]** then we need to adjust and need to add real-world training

**[00:12:20 – 00:12:23]** data, as it becomes essential.

**[00:12:24 – 00:12:26]** Lesson number two, reusable skills matter more than

**[00:12:26 – 00:12:31]** individual robots, and this is very important, that skills

**[00:12:31 – 00:12:33]** become the key differentiator.

**[00:12:33 – 00:12:36]** Skills will have the main impact in future, and only companies that

**[00:12:37 – 00:12:40]** invest in skills, not in individual robot generations, that not bet

**[00:12:40 – 00:12:43]** on individual robots, will succeed.

**[00:12:43 – 00:12:47]** And lesson number three, physical AI requires more than just

**[00:12:47 – 00:12:48]** technical aspects.

**[00:12:48 – 00:12:51]** If you think about bringing physical AI to the shop floor,

**[00:12:51 – 00:12:54]** to industrial environments, you will quickly realize it's more

**[00:12:54 – 00:12:55]** than just a technical challenge.

**[00:12:55 – 00:12:58]** It's also about data management, it's about safety, it's

**[00:12:58 – 00:13:01]** about operational integration, change management.

**[00:13:01 – 00:13:04]** Only companies who see physical AI as a whole ecosystem and

**[00:13:04 – 00:13:08]** as a whole, not only as a technical challenge, will succeed.

**[00:13:08 – 00:13:10]** And with those three lessons, I will hand over back to

**[00:13:10 – 00:13:11]** you, Michelangelo.

**[00:13:11 – 00:13:13]** Thank you so much, Nicholas.

**[00:13:13 – 00:13:16]** Alright, so with this last slide, I would like to share

**[00:13:16 – 00:13:20]** with you that our ambition is not only to understand why

**[00:13:20 – 00:13:24]** we do it, what we do it, and what the value is, but even more, the

**[00:13:24 – 00:13:28]** goal is to make something scalable.

**[00:13:28 – 00:13:32]** As I mentioned earlier, the demographic change is here to stay.

**[00:13:32 – 00:13:33]** It's not going to get any easier.

**[00:13:33 – 00:13:35]** It's not going to get any better.

**[00:13:35 – 00:13:39]** So, what we have identified is an approach that we are trying right

**[00:13:39 – 00:13:43]** now, is not only proving the value of the technology, but then to

**[00:13:43 – 00:13:47]** replicate in the production plant, in the real environment, with real

**[00:13:47 – 00:13:50]** physical efforts, with integration efforts, with everything that needs

**[00:13:50 – 00:13:53]** to happen in the real production plants, with a lot of other

**[00:13:53 – 00:13:55]** aspects that we need to consider.

**[00:13:55 – 00:13:57]** Once that has been done.

**[00:13:57 – 00:13:59]** The scalability is here to stay.

**[00:13:59 – 00:14:04]** So, we are here to not just do technology for technology.

**[00:14:04 – 00:14:06]** It's the real, do it in a way that we can scale it as

**[00:14:06 – 00:14:11]** a standard across operations where the need is, where we see

**[00:14:11 – 00:14:15]** that this is exactly what addresses it and the transformation that

**[00:14:15 – 00:14:18]** will help us to stay competitive, but to deliver the products

**[00:14:18 – 00:14:20]** to our customers and patients.

**[00:14:20 – 00:14:24]** With that, I'd like to thank everyone and the incredible

**[00:14:24 – 00:14:26]** team that stands behind this incredible work, only within

**[00:14:26 – 00:14:31]** a few months only, to EY, Bundlebots and Grid Dynamics,

**[00:14:31 – 00:14:34]** that make this even possible, to be a true partner on that, and

**[00:14:34 – 00:14:37]** to all of you for listening to us, and happy to take any questions.

