Alexandre Embry heads the AI Robotics and Experiences laboratory at Capgemini, an entity that focuses on robotics coupled with artificial intelligence. Polyfunctional and autonomous robots today constitute one of their main areas of research.
JDN. Why did Capgemini decide to make physical AI and humanoid robotics a strategic focus?
Alexandre Embry. We consider this to be a continuation of the LLM. These robots are, basically, AI agents to which we have added action capabilities in the real world: arms, legs, sensors. Their design and training are based on approaches very close to those used for software agents, with similar data models. This places us in a fairly natural position on this subject, given our initial expertise in AI and software, as well as the robotics skills we had already developed.
How does Capgemini’s expertise position you with regard to the integration of robots in companies?
If we look at the current evolution of robotics, it corresponds precisely to our skills: software development, artificial intelligence and consulting. This includes defining use cases, identifying business value, as well as supporting deployments. We work to integrate these machines into operational environments, linked to companies’ IT, IoT and OT systems.
You recently published a report on physical AI and humanoid robotics. What are the main lessons?
More than two-thirds of the business leaders surveyed (nearly 1,700 in total, spread across a wide variety of sectors) believe that physical AI will be a game changer for their industry. And 64% see it as a lever to improve their competitiveness in the years to come. These figures demonstrate significant adoption potential. On humanoid robotics, two thirds of managers believe that this will help transform their sector of activity. And 78% say they are ready to deploy them on a scale, but not immediately. Finally, 72% cite their versatility as the main incentive to deploy them.
Which sectors have the most expectations regarding these technologies?
These are essentially the sectors where collaboration between humans and robots is possible, whether to alleviate labor shortages or to protect workers working in difficult conditions. We therefore find the fairly classic sectors: manufacturing industry in the broad sense, automobile, construction, aerospace and defense, warehousing and logistics.
What deployment models are emerging?
The humanoid form is not the most cited in the short-term deployment prospects. Autonomous mobile robots, AMRs and intelligent AGVs, come first, followed by robotic arms. The humanoid form is cited by less than 25% of respondents as being likely to be deployed within three years. That said, the appetite for this form is actually enormous, because it is the most easily adaptable in environments already operated by humans. But it is also potentially the most challenging to deploy.
For what ?
The first reason is technological. The roadmap evolves from VLA (Vision-Language-Action models) towards Robotics Foundation Models, which will allow robots to increasingly adapt to heterogeneous environments, and this is where we will tend towards true general autonomy. There are still three to five years to go before we get there. The second dimension concerns operational deployment issues. This involves in particular the constraints to be anticipated: security, cybersecurity and regulation. This work must be carried out jointly with governments, because these robots are autonomous vehicles, machine tools and AI systems all in one. In Europe, the AI Act applies in this respect and must be fully integrated. Finally, there is acceptance by operators already present in these environments, an important societal subject that must be addressed.
Are costs also a barrier?
There are two separate topics. The first is the hardware. A fierce competition is taking place between China, Europe and the United States. Chinese manufacturers are today ahead in their ability to deliver at scale, but they will be caught up, while American players are picking up the pace. All this will automatically lower prices.
And the second subject?
This is the time required to train a robot to execute a given scenario. The challenge is to arrive at a standardizable framework, what we call a “training factory”: industrialize the training itself, with a fixed cost per scenario, which decreases as the computing power and the models progress. Because these robots cannot be programmed: they train on data. And the particularity here is that we operate in the real world (point clouds, real-time video analysis, 3D environment capture) which is much more complex than traditional textual or video data.
What economic models will prevail in humanoid robotics?
Several models will coexist, as with all industrial machines. Chinese suppliers sell pure hardware, a model that has its place. In the United States, we are seeing the emergence of pay-per-use. And there are the classics of leasing. What is unlikely, however, is the hardware delivered alone: there are too many additional services required (platform, computing, maintenance, updates) for this model to hold up. Players will therefore position themselves as all-inclusive robotics-as-a-service. These models will remain linked to geography: American pay-per-use, anchored on the value of substituted work, will be less relevant in Europe where underemployment remains significant. Rather, we will see a standardized robotics-as-a-service, packaging all costs in OPEX.
How can the French humanoid robotics ecosystem stand out?
We don’t start from a blank sheet. I am thinking in particular of what was recently announced in the automotive sector, a model that goes beyond a simple partnership, with real collaboration on the manufacturing of robots. The main obstacle remains financing: we do not have the same investment mindset in Europe as in the United States, nor the massive state support that we observe in China. This is why industrial partnerships are particularly relevant. But we are not helpless. And regulation, often seen as a hindrance, can become an advantage: Chinese robots tick relatively few boxes for European industrial deployment. Our ability to provide a framework guaranteeing compliant and sovereign robots represents a real competitive advantage, exportable beyond Europe. Our card is not to try to compete on the mass market, but to target the specific needs of European industry and to support a few key players.
