Comau brings NVIDIA computing and simulation into industrial robotics

Comau is integrating NVIDIA edge computing and simulation technology into its industrial robotics portfolio, combining new controller architectures with virtual development of robot behaviour and vision-based applications. The programme includes Jetson computing for the next generation of industrial robot controllers, Isaac Sim models of Comau robots and training workflows for palletising and depalletising equipment.

Announced in Turin on 21 October, the collaboration brings NVIDIA’s accelerated computing and physical AI tools into Comau’s product development and system engineering processes. The work covers industrial, collaborative and mobile robots, with the company preparing reusable simulation models and exploring AI control strategies before their application to physical equipment.

The next generation of Comau industrial robot controllers will incorporate NVIDIA Jetson edge computing technology. These processors provide resources for AI inference and other parallel workloads, allowing computationally demanding applications to operate within or near the equipment they support.

Industrial robot control systems coordinate motors, sensors and programmed movements within defined mechanical limits. Their behaviour depends on accurate information about position, speed and the operating state of the equipment, while communication with surrounding machinery must remain consistent with the requirements of the production process.

Traditional motion-control functions commonly operate according to predictable timing requirements. In contrast, AI perception can impose varying processing demands as the scene changes or additional objects are detected. Combining those workloads within one controller architecture requires attention to execution timing, memory allocation and the communication paths used by the different functions.

Jetson provides accelerated processing that can support camera analysis and learned models, but its presence does not independently establish that every AI application will meet the real-time requirements of a particular robot. The complete controller must allocate resources and verify the timing and behaviour of its control functions under representative operating conditions.

Comau has not disclosed the specific Jetson processor models or completed commercial controller configurations covered by the programme. Detailed computing performance, electrical consumption and the availability of individual controllers will therefore depend on subsequent product information.

The collaboration also involves the integration of Comau’s industrial, collaborative and mobile robot portfolio into NVIDIA Isaac Sim. The simulation platform allows developers to construct virtual environments in which robot geometry, surrounding equipment and selected physical interactions can be examined before installing a complete workcell.

Comau is contributing SimReady robot assets using OpenUSD to the SimReady Central Library hosted on Hugging Face. These assets provide reusable representations of equipment that can be incorporated into different simulation arrangements, reducing the need to reconstruct the same robot model independently for every development project.

OpenUSD provides a structured method for describing three-dimensional scenes and their components. Its use can support the exchange of models between compatible applications, although the engineering value of an asset depends on the physical information included alongside its visual representation.

A model used for robot control development must represent more than the appearance of the machine. Joint positions, motion limits, geometry and relevant dynamic characteristics influence whether simulated movements resemble those of the physical equipment.

The accuracy of those properties matters when a robot interacts with other objects. Contact forces, friction, object deformation and actuator response can all influence whether an attempted manipulation succeeds. A visually convincing simulation may produce misleading control results if the relevant physical behaviour has not been represented adequately.

Comau is using NVIDIA Isaac Sim together with Isaac Lab and discrete-event simulation technologies to develop AI applications for intralogistics. The company’s MI.RA machine vision systems are among the areas where these tools will support training and investigation of control strategies.

One focus is palletising and depalletising, in which a robot must identify objects, decide how to handle them and move them into or out of a structured arrangement. The task can become demanding when the objects differ in size, orientation or packaging, or when the available space changes as material is added or removed.

Vision-based perception provides information about the location and apparent geometry of an object. The controller must then select an appropriate movement within the robot’s mechanical limits and coordinate the gripper with the intended contact points.

Reinforcement learning can be used to develop such behaviour by repeatedly evaluating actions in a simulated environment. A policy may be adjusted according to whether its actions achieve the defined objective, allowing the development process to explore a range of possible movements without repeatedly exposing physical equipment to unsuccessful attempts.

Isaac Lab provides tools for organising those training processes, while Isaac Sim contributes the virtual environment and associated physical modelling. Comau is also using discrete-event simulation to examine the progression of activities through wider logistics workflows.

These simulation approaches answer different engineering questions. Physics-based simulation can help determine whether a robot can perform an individual movement or handling operation, while discrete-event models can examine how the duration and sequence of those operations affect the performance of a wider material-handling process.

A robot may execute a particular pick successfully yet still create a bottleneck if its cycle time does not meet the requirements of the surrounding equipment. Conversely, a logistics model may predict an acceptable process rate without representing the detailed contact conditions that determine whether every object can actually be handled.

Combining the models provides a way to investigate both aspects of the application, although the resulting predictions remain dependent on the data and assumptions used. Relevant equipment speeds, processing intervals, object characteristics and operating constraints must be represented accurately if the findings are to support engineering decisions.

Comau intends to train and evaluate AI-driven control strategies in these environments before transferring them towards physical applications. The approach may reduce the amount of development work requiring a complete installed workcell, particularly where many variations of an object or process must be investigated.

However, simulation cannot reproduce every condition encountered in a factory or distribution centre. Sensor noise, calibration variation, mechanical wear, unexpected obstructions and differences in object behaviour can all cause a physical system to respond differently from its virtual counterpart.

Application validation must therefore continue after simulation, including checking the robot’s behaviour within the actual operating environment. Learned actions must remain compatible with the controls and protective measures required by the installed machine.

Functional safety remains a separate engineering requirement. Industrial robots operate within defined safety arrangements, and introducing AI-generated movements does not remove the need for appropriate risk assessment, protective functions and verification of the complete application.

Comau’s collaboration is intended to connect hardware development, simulation and application training, but the company has not reported an independently measured improvement in deployment time, production throughput or manipulation accuracy from the new integrations.

The current programme establishes specific development work involving Jetson-based controllers, reusable robot models and AI training tools. The practical benefits will depend on the behaviour of finished controller products and the results of application testing under the mechanical, electrical and operational conditions encountered by customers.