"We don't run apt upgrade anymore." ... wait, what?
Robotics engineers aren't being lazy. They're scared. And honestly? They're right to be.
Here's how it happens.
You start on Debian, Ubuntu, or Pi OS — of course you do. It's what you know, it boots in minutes, and you can pip install your way to a working demo by lunch.
Then the demo becomes a product.
You ship 20 devices. Then 200. Then you're staring down 2,000.
And somewhere along the way you ran one apt upgrade that swapped a kernel module, bumped a CUDA dependency, or quietly broke a camera driver — and you burned three days bisecting it on a device you can't physically reach.
So you make the rational call: stop upgrading. Freeze everything. Don't breathe on it.
Except...
→ You can't apply a CVE patch without risking the whole fleet
→ Every device has drifted into its own little snowflake
→ There's no clean rollback when a remote update half-applies and bricks
→ And "reproducible" means a 40-step readme that only the engineer who wrote it can actually run
Docker helps. It genuinely does. But the container is the isolated part, the host underneath isn't. That mutable Debian base is still sitting there, one upgrade away from pulling the floor out from under you.
None of this shows up in the lab. It shows up in the field. At scale. Usually on a Friday.
And the fix, when teams finally admit they need one, is a 3-month table flip onto a proper embedded foundation... right when they should be shipping features.
Here's what I keep coming back to:
These teams didn't make bad calls. They made the easy call, because the easy on-ramp pointed straight at Debian. The cost just got deferred to the worst possible moment.
Zephyr just turned 10. Worth paying attention to why.
In 2016 the microcontroller world had no default OS. Every team rolled its own RTOS or bought a proprietary one.
Zephyr replaced that with a productized layer: one OS, maintained, supported for the long haul. Teams stopped building the plumbing themselves.
A decade later it is the default at that tier: 1,000+ supported boards, 70% of surveyed North American orgs shipping it in production, 69% planning to increase adoption this year.
The lesson is not "RTOS is hot." It is that embedded teams stop rolling their own OS the moment a productized layer does the enablement for them.
That layer does not exist one tier up. '
At the microprocessor and edge GPU level the OS is already Linux, built with Yocto. But there is no standard.
Every team still hand-builds a bespoke image per board, maintains it manually, and redoes the work when the silicon changes. And the enablement gap there is not a BSP problem.
A better BSP is only one slice.
The actual work is:
- toolchain mastery most teams do not have on staff
- dedicated build compute to stand up and maintain
- 1:1 vendor-layer integration, per silicon, redone every time it changes
This is the exact condition the MCU world was in before Zephyr. Same fragmentation, same lock-in, same maintenance tax.
The difference is that this tier is where physical AI compute is landing, so the prize is larger.
That is the gap Avocado OS closes. We do the toolchain mastery, the build infrastructure, and the deep vendor-layer integration, and deliver a productized layer across 16+ targets.
Built on Yocto, OSS, and the NVIDIA and NXP partner programs. Zephyr proved teams want this one floor down. The tier with the GPUs on it has a harder gap and a bigger prize.
Avocado OS is built to be the standard there.
https://t.co/e54ZUANkq4
@Stryker_Eureka0 Have been evaluating support for this on https://t.co/moACdfO9Tw. It’s a really interesting product (have one in office). Would love to hear more where you’re seeing this get used
the best part of @perplexity_ai computer is that they're obviously using it to manage accounts receivable and and billing
been trying to cancel a $350/mo subscription since i signed up, never worked... please ya'll come on