At the Orano Melox site in southern France, a humanoid robot named Hoxo (short for Humanoid melOX Orano) has quietly begun a new chapter in industrial history.

Currently being tested at the nuclear fuel manufacturer’s training school, it moves, perceives, and acts in a way that looks strikingly human, yet it is powered by something profoundly new. Hoxo is not programmed; it is trained.

Behind this achievement stands the collaboration between Orano and Capgemini. Together, they combine industrial expertise, applied AI innovation, and advanced computing technologies to create intelligent, adaptable, and autonomous physical agents capable of working safely alongside humans in demanding environments.

From programming to learning

For decades, robots have been rigid by design. They excelled at repetitive, pre-programmed tasks but struggled with anything that required flexibility or judgment. Physical AI changes this paradigm.

Instead of encoding every motion and instruction through manual programming, humanoid robots now learn by experience. They absorb data from simulated environments, human demonstrations, and real-world interactions. Over time, they develop a form of embodied intelligence: the ability to perceive, reason, and act in dynamic settings.

At Capgemini’s AI Robotics & Experiences Lab, engineers have developed a training pipeline that integrates NVIDIA Isaac Sim and Isaac Lab, open source robotics simulation and learning frameworks  with advanced Vision-Language-Action (VLA) models. In this environment, Hoxo’s digital twin learns thousands of variations of a task, opening a valve, climbing a ladder, handling a tool, within safe, photorealistic virtual replicas of Orano’s facilities. Once trained in simulation, the robot transfers these skills to the real world with high precision.

This “Sim-to-Real” approach removes one of robotics’ biggest constraints: the need for extensive reprogramming for every new application. Each new mission becomes a learning opportunity, not a coding exercise. As a result, the robot becomes a versatile platform, easy to retrain and repurpose.

Vision, language and action: a unified intelligence

At the heart of this new generation of humanoids lies NVIDIA Isaac GR00T N, a Vision-Language-Action framework that merges visual understanding, linguistic reasoning, and motor control within a single AI system. Orano and Capgemini are extensively evaluating the GR00T N model as part of their Proof of Concept (PoC) for their Hoxo humanoid.

Through its cameras and sensors, Hoxo perceives its environment, identifies objects, reads labels, and understands spatial relationships. When an operator provides a command such as “Inspect the containment area and report anomalies” the robot translates that instruction into a series of physical actions.

This integration allows it to adapt to changing conditions. If an obstacle appears, it adjusts its path. If the lighting changes or the workspace evolves, it recalibrates its perception in real time. The result is a robot that continuously links perception, understanding, and execution.

NVIDIA Isaac GR00T N dual-system architecture, inspired by human cognition, makes this possible. A high-level reasoning layer interprets goals and scenes, while a fast, low-level control layer converts them into smooth, precise movements. This combination gives humanoid robots the responsiveness and generalization they need to operate autonomously in complex industrial environments like Orano’s.

Teaching by teletraining

Not all learning happens in simulation. Another method developed by Capgemini is teletraining, where human operators remotely demonstrate actions that the robot observes and imitates.

Using advanced sensors and computer vision, the robot analyzes gestures, trajectories, and timing, refining its movements through reinforcement learning. Over time, it becomes more autonomous and requires less supervision. This approach makes it possible to transfer human expertise directly to robotic systems, preserving operational know-how and amplifying it across multiple sites.

Why the humanoid form matters

In industries built by humans for humans, form is function. The humanoid design allows robots like Hoxo to navigate existing facilities naturally—opening doors, handling tools, operating panels, without the need for redesigning infrastructure.

In the future, robots could enter hazardous zones in Orano’s plants, perform repetitive or physically demanding tasks, and assist teams in maintaining productivity under challenging conditions. Their flexibility allows them to switch roles, move from inspection to logistics, or maintenance depending on operational needs.

Towards collaborative physical AI

What truly sets these robots apart is their integration into a larger ecosystem of intelligent agents. They can exchange information with digital twins, interact with virtual AI systems, and collaborate with human operators in real time.

This convergence of physical and digital intelligence defines the essence of physical AI. It is not automation in isolation but collaboration at scale.

  • For Orano, this represents a new phase in industrial innovation: safer operations, greater agility, and enhanced performance.
  • For Capgemini, it brings to life the vision of intelligent industry, where AI extends from data centers to factory floors.
  • For NVIDIA, it showcases the power of its AI infrastructure, open models, libraries and frameworks necessary to develop the next generation of AI-driven robots.

The last mile of industrial automation

Physical AI is the final link between digital intelligence and real-world execution. These humanoid robots are not static machines; they are adaptive systems that learn, reason, and evolve.

As models advance and simulations become more realistic, generalist robots like Hoxo will be able to handle an ever-wider range of industrial tasks. They will contribute directly to operational resilience, efficiency, and safety.

The collaboration between Orano and Capgemini with NVIDIA technology demonstrates what happens when industrial expertise meets advanced AI and computing: machines that no longer need to be told exactly what to do, because they understand what needs to be done. This is the future of robotics, already unfolding on the factory floor.