Raspberry Pi vs. NVIDIA Jetson vs. Intel NUC: Choosing the Right Brain for Your Robot

TL;DR: When you compare Raspberry Pi vs Jetson robot computing, the choice comes down to workload, power budget, and software stack. Raspberry Pi 5 fits teleoperated rovers, sensor bridging, and lightweight ROS 2 nodes at 5-15 W. NVIDIA Jetson Orin modules dominate when you need on-board CUDA, deep learning inference, or visual SLAM, delivering up to 275 TOPS on AGX Orin. Intel NUC (and similar x86 mini PCs) wins when you need full x86 compatibility, fast single-thread performance, and PCIe expansion for industrial sensors. Leo Rover ships with Raspberry Pi 4 as the default brain and supports Jetson or NUC payloads when workloads grow.

Why the choice of on-board computer defines your robot

The compute platform sets the ceiling for everything else on a mobile robot: sensor bandwidth, perception latency, autonomy stack, and battery life. Picking the wrong board means either burning watt-hours on idle silicon or starving your ROS 2 graph during navigation. The Raspberry Pi vs Jetson robot decision, extended with x86 mini PCs like the Intel NUC, is the most frequent architectural question Fictionlab’s team hears from researchers and integrators working with the Leo Rover.

This article breaks down the three families by real specifications, ROS 2 maturity, thermal behavior in field robotics, and the workloads they actually handle. No fanboyism, no marketing claims – just what each board does and where it stops being the right tool.

What Raspberry Pi offers a mobile robot

The Raspberry Pi 4 and Raspberry Pi 5 are ARM64 single-board computers built around Broadcom SoCs. The Pi 5 uses a quad-core Cortex-A76 at 2.4 GHz with up to 16 GB LPDDR4X. There is no dedicated AI accelerator, though the VideoCore VII GPU handles graphics and hardware HEVC decode, and the optional AI Kit adds a Hailo-8L module rated at 13 TOPS.

For ROS 2 (Humble and Jazzy LTS on Ubuntu 22.04/24.04 ARM64, plus newer releases where packages are available), the Pi is a first-class citizen. You can run navigation2, micro-ROS bridges, robot_state_publisher, teleop, and a modest perception pipeline. The Leo Rover uses a Raspberry Pi 4 as the default compute unit, exposing GPIO, CSI cameras, and USB to ROS 2 packages maintained by Fictionlab.

Where the Pi stops being enough is dense perception: real-time stereo depth at 30 fps, multi-camera visual SLAM, or any transformer-based vision model. The CPU saturates, thermal throttling kicks in above approximately 80 °C without active cooling, and you lose determinism in your control loop.

When to pick a Raspberry Pi

The Pi is the right brain when your robot does the following:

• Teleoperation with one or two cameras and a joystick or web UI
• Differential drive control via micro-ROS on an MCU, with ROS 2 running navigation and mapping using 2D LiDAR
• Educational and research projects where ARM64 ROS 2 packages cover the use case
• Power-constrained UGVs targeting 5-10 W idle, 15 W peak compute draw

What NVIDIA Jetson brings to autonomous robotics

The Jetson family is NVIDIA’s embedded line that pairs ARM Cortex CPUs with CUDA-capable GPUs and, on Orin, dedicated deep learning accelerators (NVDLA). The current relevant modules for mobile robots are Jetson Orin Nano (up to 67 TOPS sparse INT8 on the 8 GB Super configuration, per NVIDIA published specs), Orin NX (up to 157 TOPS), and AGX Orin (up to 275 TOPS).

JetPack 6 ships Ubuntu 22.04 with CUDA 12, TensorRT, cuDNN, and DeepStream. ROS 2 Humble runs natively, and Isaac ROS provides GPU-accelerated nodes for visual SLAM (cuVSLAM), Nvblox 3D reconstruction, and image processing. For any robot that needs to run YOLO, segmentation, or point cloud inference on board, Jetson is the path of least resistance.

The trade-offs are real. Orin Nano draws 7-25 W, Orin NX 10-25 W, and Orin AGX 15-60 W depending on power mode. Carrier boards add cost and complexity. JetPack often lags upstream Ubuntu and ROS releases, and some Python ML packages need rebuilds for ARM64 + CUDA.

When to pick a Jetson

Jetson is the right brain when your robot does the following:

• On-board deep learning inference (object detection, semantic segmentation, pose estimation)
• Visual or visual-inertial SLAM with stereo or RGB-D cameras at 30+ fps
• 3D reconstruction, occupancy mapping with Nvblox, or LiDAR + camera fusion
• Outdoor autonomy stacks for research platforms running Isaac ROS or custom CUDA kernels

What Intel NUC and x86 mini PCs deliver

Intel NUC (now continued by ASUS) and equivalent x86 mini PCs (Beelink, Minisforum, industrial fanless units) run standard Ubuntu, Windows, or real-time Linux kernels without ARM porting headaches. A modern NUC with Core i5 or i7 (13th/14th gen, or Ultra series) gives you strong single-thread performance, AVX2, fast NVMe storage, Thunderbolt, and 2.5 GbE.

For ROS 2, x86 is the reference platform. Most packages are released and tested first on amd64. If your stack depends on Open3D with specific wheels, PyTorch with CUDA on a discrete GPU, or proprietary sensor SDKs (Ouster, Velodyne, ZED, Basler), x86 removes friction. Some NUC variants also support an external GPU over Thunderbolt or USB4, though that is unusual in mobile robotics due to power and vibration.

Power consumption is the catch. A Core i5 NUC typically draws 8-15 W idle and 35-65 W under load. You need a 19-20 V supply, which means a DC-DC converter on most battery-powered rovers. Thermal design is also more demanding than a passively cooled Pi.

When to pick a NUC or x86 mini PC

An x86 mini PC is the right brain when your robot does the following:

• Runs proprietary sensor or simulation software that ships only for amd64
• Needs heavy CPU multi-threading (motion planning with OMPL, large factor graphs in GTSAM, dense ICP)
• Acts as a development twin where you want identical binaries on workstation and robot
• Has the power and space budget for 30-60 W of compute, typical of larger UGVs in space analog and field research deployments

How the three platforms compare on specs and ROS 2 support

The table below summarizes the boards most relevant to mobile robotics today. Prices are approximate street prices for the module, development kit, or barebone unit as noted and exclude carrier boards, storage, and RAM where applicable.

Numbers come from Raspberry Pi Ltd, NVIDIA Jetson datasheets, and Intel ARK product pages. Power figures are vendor-rated TDP or measured envelopes from published reviews; treat them as approximate operating ranges.

How to match the platform to the workload

The Raspberry Pi vs Jetson robot question, plus the NUC option, becomes simpler when you map workload to silicon. The matrix below is a starting point Fictionlab’s team uses with customers evaluating compute for the Leo Rover and larger platforms.

• Teleop + 2D LiDAR SLAM (slam_toolbox, Cartographer): Raspberry Pi 4 or 5 is sufficient.
• Single RGB-D camera, classical perception, Nav2: Raspberry Pi 5 with active cooling, or Jetson Orin Nano if you want headroom.
• YOLOv8 / YOLO11 at 15-30 fps on board: Jetson Orin Nano (small models) or Orin NX (medium/large).
• Visual-inertial SLAM, Nvblox, multi-camera fusion: Jetson Orin NX or AGX Orin.
• Heavy planning, simulation co-execution, proprietary SDKs: Intel NUC i5/i7 or Core Ultra.
• Hybrid (perception on Jetson, planning on NUC): Two-computer architecture connected over Gigabit Ethernet, common on larger UGVs.

What powers the Leo Rover today

The Leo Rover ships with a Raspberry Pi 4 (4 GB) as the default compute unit. That choice reflects the rover’s role as an open-source research and education platform: low idle power (the rover runs 6+ hours on its internal battery), full ROS 2 support, and easy access to GPIO and CSI for custom payloads. The Pi handles teleoperation, the on-board camera stream, micro-ROS communication with the motor controller, and basic Nav2 with a 2D LiDAR add-on.

For users who need more, the Leo Rover’s payload bay and 5 V / 12 V rails accept a Jetson Orin Nano or an Intel NUC mounted alongside the Pi with suitable DC-DC power conversion where needed. A common pattern is to keep the Pi as the low-level bridge to motors and sensors, and add a Jetson for perception. This split keeps the low-level motor loop isolated while giving you CUDA where you need it.

FAQ

Is Raspberry Pi 5 enough for ROS 2 Nav2 with a depth camera?

For low-resolution depth (424×240 or 640×480 at 15 fps) and Nav2 with a 2D costmap, yes, with active cooling. For 720p or 1080p depth at 30 fps plus 3D costmaps, you will saturate the CPU and should consider a Jetson Orin Nano.

Can you run CUDA workloads on a Raspberry Pi?

No. The Raspberry Pi GPU is not CUDA-capable. If you need CUDA, you need a Jetson or an x86 machine with an NVIDIA dGPU. The Pi 5 with the AI Kit (Hailo-8L) covers some inference workloads but uses a different SDK.

How does Jetson Orin Nano compare to Intel NUC for SLAM?

For visual SLAM with Isaac ROS cuVSLAM or CUDA-accelerated SLAM pipelines, Orin Nano typically wins on frames per watt. For CPU-bound SLAM (RTAB-Map default, Cartographer, standard ORB-SLAM3), a Core i5 NUC delivers higher single-thread throughput. Match the algorithm to the silicon, not the other way around.

What about power consumption in the field?

Approximate field measurements: Raspberry Pi 5 runs 5-10 W under typical ROS 2 load, Jetson Orin Nano 10-15 W with perception active, Intel NUC i5 25-40 W under sustained load. On a 100 Wh battery, that is roughly 10-20 h, 7-10 h, and 2.5-4 h of compute-only runtime before motors and sensors.

Which platform has the best long-term ROS 2 support?

All three have active support. x86 is the ROS 2 reference platform and gets new distros first. Ubuntu ARM64 is a Tier 1 ROS 2 platform for current LTS releases. Jetson tracks ROS 2 through JetPack, which lags behind upstream Ubuntu and ROS releases but adds Isaac ROS packages you cannot get elsewhere.

Can you mix platforms on one robot?

Yes, and it is common on larger UGVs. A typical split is a microcontroller handling real-time motor control, often with a Raspberry Pi for sensor bridging, with a Jetson or NUC running perception and planning. ROS 2 DDS handles the inter-node communication over Ethernet.

Does the Leo Rover support Jetson out of the box?

The Leo Rover’s payload bay, power rails, and Ethernet expose what a Jetson needs with the appropriate carrier and power adapter. Fictionlab’s documentation covers adding a Jetson alongside the default Raspberry Pi for users building perception-heavy applications.

Choosing the right brain is a workload decision, not a brand decision. Start with the algorithms you must run, measure the power envelope you can afford, and pick the smallest board that clears both bars. If you are sizing compute for a Leo Rover or a custom UGV based on Fictionlab hardware, browse the available compute and payload options in the shop to match a configuration to your project.