Three jobs, three rhythms
A robot’s computer does three things with different urgencies. Motor control cannot tolerate delays of milliseconds. Perception, which turns cameras and LiDAR into a map, works tens of times per second. And reasoning, which understands an instruction or plans a task, can take a second.
The first always goes on the robot. The third can go elsewhere. The second is where the design is decided.
On board or on a server
Carrying the computing on board gives autonomy: the robot keeps working if it loses the network. In exchange, it weighs, drains the battery and heats up. Leaving it on a server saves all of that, but the robot depends on coverage and on the latency of the wireless network.
The usual split is mixed: control and perception on board, large models and training on a server with graphics cards, and a degraded mode for when there is no network.
Power and consumption: the battery rules
A robot supplies 24 or 48 V DC from its battery, with dips when the motors start. The computer needs a wide-range converter and a clean shutdown when the battery runs out.
Every watt the computer uses is autonomy the robot loses. That is why unified memory computers fit: they run vision and language models with far lower consumption than a dedicated desktop graphics card.
Connections and real time
Count the sensors: cameras over Ethernet or USB 3, LiDAR over Ethernet, and the motors over CAN or EtherCAT. Ask for the ports as standard, not adapters dangling inside the chassis. If control runs on the same computer, a Linux kernel prepared for real time is needed.
ROS 2 runs on Ubuntu on x86 and ARM processors. Choosing x86 gives access to all PC software and makes it easy to develop on the desk exactly what later goes in the robot.
NVIDIA graphics or unified memory
Many accelerated perception libraries, such as NVIDIA’s for ROS, require an NVIDIA graphics card. If your robot depends on them, the compact must carry an RTX PRO. If what you need is for the robot to understand instructions and describe what it sees with a large model, unified memory gives more memory for less power.
| Case | Computer | Starting point |
|---|---|---|
| Mobile robot with navigation and classic vision | Low-power compact on board | Efficient processor, 32 GB, ports for LiDAR and cameras |
| Robot that understands instructions and describes what it sees | Unified memory compact on board | 64 GB of unified memory, powered from the battery |
| Accelerated perception with NVIDIA libraries | Compact with professional graphics | RTX PRO with 16 to 24 GB in an enclosure of about 10 litres |
| Training and simulation | Server or workstation with graphics cards | Several graphics cards and plenty of memory, off the robot |
From the prototype to the fleet
When the robot goes from one to twenty, the computer has to be identical in all of them: same board, same firmware version and same image. We build it in series with your cloned image and a serial number per unit, and we warn you in advance if a part is going to be discontinued.
Related hardware and guides
Frequently asked questions
x86 or ARM for a robot with ROS 2?
ROS 2 runs on both. ARM uses less power in small devices. x86 gives more performance, more memory and the same software you use when developing, and accepts professional graphics cards. For robots with large models on board, x86 with unified memory is a good balance.
Can it run from the robot’s battery?
Yes, with a DC converter for 24 or 48 V that withstands the dips of the motors starting. It is sized for each project.
Do I need an NVIDIA graphics card?
Only if you use libraries that require it, such as NVIDIA’s accelerated perception ones, or if you train models. To run open language and vision models, unified memory is enough and uses less power.
And training the models?
Off the robot, on a server or a workstation with several graphics cards. The robot runs the trained model. We propose both machines together so they share the system and versions.
What will you use it for?
Tell us which models you want to run, how many people will use it and where it will work. We will propose the compact, with or without a professional graphics card, and its price within 12 working hours.