NVIDIA Jetson

NVIDIA Jetson is a family of compute modules that serve as the "brain" in many robots: low-power computers with a graphics processor and accelerators for neural networks. NVIDIA does not build robots; it supplies a board with the JetPack system and tools on which integrators run perception, planning and AI models.

Of the documents we hold, the Jetson AGX Orin series is described best. The technical brief from July 2022 states that the 64 GB and 32 GB modules deliver up to 275 TOPS (64 GB version) and 200 TOPS (32 GB version) in INT8 arithmetic, have a 12-core CPU (64 GB version), 32 or 64 GB of LPDDR5 memory with 204 GB/s of bandwidth, and power profiles from 15 to 60 W. The maker describes this as server-class performance for robots and autonomous machines: multi-sensor fusion, 3D perception, language understanding. The younger, cheaper Orin NX and the newer Thor generation appear in specifications in our robot catalogue, although we do not hold their datasheets.

The list of robots with a Jetson in our catalog is long. Atlas has, according to its description, a Jetson Thor T5000 computer, and Booster T2 a Thor chip rated by the maker at 2,070 TFLOPS, while the earlier Booster T1 runs on an AGX Orin of up to 200 TOPS. Booster K1 has an Orin NX. Unitree G1 EDU uses an Orin and Unitree Go2-W an Orin NX 16 GB (100 TOPS). It is also found in Agibot A2 Ultra, EngineAI PM01, LimX Oli and in the mobile manipulator Franka Mobile FR3 Duo. The announced humanoid LG Bipedal Humanoid is to be built on the Thor platform.

For cost and performance this is one of the more important choices, because computing power and current draw affect runtime and cooling. A high-power module raises the cost and battery use, while a smaller one may not carry large models such as those in “VLA Models: How AI Controls Humanoid Robots in 2026”. Data from cameras, for example Intel RealSense or Luxonis, goes to just such a board. The general picture of a robot is covered in “How Does a Humanoid Robot Work”.

In Poland, Jetsons are used among others by Antmicro, which according to our materials builds Edge AI platforms for robotics based on these modules. Our materials name no specific Polish NVIDIA distributor; the modules are available through electronics distributors in Europe. Check the availability of versions and prices, because the offer changes quickly.

For a buyer, the distinction between a bare module and a developer kit also matters. A module needs a carrier board with camera inputs, power and connectors, which the robot must have of its own, while the kit serves for prototypes. In robots with a Jetson, the cost of the board, cooling and integration therefore adds up, as does software support on the robot maker's side. For a reader choosing a robot, more important than the TOPS figure itself is the question of what the robot can do without the cloud: whether it recognises objects and plans motion locally or needs a constant link. “How Much Does a Humanoid Robot Cost in 2026” helps here too.