Jetsons seemed to me to have the most mature ecosystem, but some of those price increases surely puts them away from reach of students/hobbyists, and can have also impacts on larger fleet deployments.
Luckily I just bought an Jetson Orin Nano for the old price
I’m not using anything production yet, but I’ve recently researched exactly this question, so these are the options besides the ones you mentioned:
If your algorithms need CUDA, or you do not want/can’t invest in rewriting them for a different execution backend, there is really no alternative to Nvidia GPUs. You could buy a cheap desktop/laptop GPUs and get them to run on a Raspberry Pi via PCIe. This was previously reported as generally working. The Jetson Thor has a Blackwell-architecture GPU with 2560 cores, so it is comparable to an RTX 5050. Mechanically speaking, you could use MXM modules normally used in laptops. So this is the DIY-Jetson solution that gives you Jetson-Thor-level GPU compute for a much lower price (an RTX 5050 costs ~350$).
You can use mobile SoCs and run the neural networks on the mobile GPU (this is what openpilot does). Most support only Android-based drivers (the drivers themselves are Linux-drivers, but they apparently require Android SDKs on top), but there is one SoC that supports Linux: The Rockchip RK3588(S). It has 8 core CPU (4 ARM A76 + 4x A55), a Mali G610 mobile GPU and interestingly, an AI accelerator (NPU) that supports FP16. For most units it already has Linux-drivers, but for some subunits, Linux-drivers are still missing, notably HDMI output and video encoding/decoding. Boards with 8GB RAM are available from 200$.
I’ve seen papers where research labs use Intel NuCs and generally mini-PCs on drones. Generally, x86-based CPUs have a higher power consumption than ARM CPUs. Also, no CUDA support, although Intel has some interesging NPUs in their most recent laptop SoC used in mini-PCs
Ultimately, I still think the Jetson is the best option if you need AI acceleration for computer vision algorithms that require FP16 precision instead of only INT8: Lowest risk regarding software support, single chip CPU + GPU. I would not buy older Jetson TX2 boards though, because they do not have software support anymore.
NXP NavQ plus for an ackermann robot. Have not tried the slam toolbox yet…but this board only has 4gb ram so 2d slam is probably the most I can run?
I have tried littleslam_ros2 but it can’t keep up with faster moving robots.
I am also using an amd kria kv260, but also again limited to 4gb. You can use the FPGA logic to do things…it’s a bit easier now due to high level synthesis but the learning curve is still steeper than running isaac ros.
The jetson orins were kind of the baseline for being able to run 3D SLAM on edge, but I am sure there are other boards that can do it. It is hard to build the 3D SLAM packages on something with < 8GB ram.
The RK3588 is great! It’s the chip used on Orange Pi 5 and it has hardware encoders for h264 video – essential when you need to remotely monitor or teleop/tele-assist your robot with live video.
I used an extremely similar product whose only difference is the RAM speeds, but I actually went into the BIOS to reduce the RAM speeds to match those of this new product so that it should be precisely the same!
