Silent
Joint signs and order
A checkpoint trained on an arm whose second joint runs the other way drives yours backwards. Nothing warns you. joint_signs = [1,-1,1,1,1,1] is the difference between working and not.
~/ The install step for robot policies
Get your robot doing real tasks tonight.
The robot brains are already free. Getting one onto your arm — the joint signs, the calibration, the normalisation stats, the camera order — is the part that quietly breaks. AVIOS is the install step.
[ ]Joint signs and orderunchecked
[ ]Normalisation statisticsunchecked
[ ]Calibration offsetsunchecked
[ ]Camera key orderunchecked
Each of these loads cleanly and emits plausible numbers. Today nothing checks them before the arm moves.
keptdiscardedepisode 0about ten minutes of teleoperation
The brains are free
Downloading a robot brain is already solved
Open weights, permissive licences, consumer hardware. None of this is the hard part any more, and AVIOS will never try to compete with it.
| Model | Licence | Runs on | Robot-specific data its docs ask for |
|---|---|---|---|
| openpiπ0 · π0-FAST · π0.5 | Apache-2.0 | 1× RTX 4090 | 30 demos |
| MolmoAct2Ai2 · weights + training data | Apache-2.0 | < 21 GiB VRAM | under 200 demos |
| SmolVLA450M · the cheap way in | Apache-2.0 | CPU / MacBook | tens of demos |
| GR00T N1.7NVIDIA · new_embodiment tag | NVIDIA Open Model | workstation GPU | “a few thousand steps” |
Click a row to inspect it. Checked October 2026. The GR00T repository describes its own weights as Apache-2.0 in places; the licence shipped with them is the NVIDIA Open Model Licence. Keeping this table honest and current is itself part of the job.
What breaks
What actually breaks
Not the model, and not a shortage of data. Four silent misconfigurations, each of which loads cleanly, emits plausible numbers, and then moves the arm wrongly.
Silent
Normalisation statistics
No open stack ships usable stats for a robot it has not seen. Compute them wrong and the policy trains to a believable loss and actuates garbage.
Silent
Calibration offsets
One miscalibrated joint produced the same failure across three different policy architectures. No amount of extra training fixes it.
Silent
Camera key order
Swap wrist and overhead and the policy still runs. It just reaches for the wrong place, confidently.
Break it yourself
Policy loss0.031
Confidence0.97
Miss0 mm
Wired correctly. The arm goes where the policy meant.
And the reason this is worth a product: an under-trained policy and a wrongly-wired one produce the identical symptom — a robot that twitches and does nothing. Today you cannot tell which problem you have.
The loop
The loop AVIOS is for
One cheap arm, one afternoon, one task that works. Then the next task costs you ten minutes instead of a weekend.
1
Describe
One versioned file for your arm: joint order, signs, limits, home pose, cameras, gripper polarity.
2
Calibrate
A scripted routine the arm runs before anything learned is allowed to move it.
3
Record
Teleoperate the task for about ten minutes. Guidance on which motions, and which episodes to throw away.
4
Train and run
Fine-tune, then run behind a watchdog with joint limits, delta clamps and a deadman.
Ten minutes of demonstrations is not optimism: ACT/ALOHA (Zhao, Kumar, Levine & Finn, 2023) hit 80–90% on six fine manipulation tasks from exactly that, on deliberately low-cost hardware.
The missing number
The number nobody has measured
How many demonstrations does your arm need, for your task, to reach a success rate you would accept? Every stack publishes a different answer and none of them publishes a curve.
30 demos
Demos you record · log scale
Success rate on your arm
—
Not measured by anyone yet. That is the curve AVIOS will publish.
openpidocumented fine-tuning example
30 demos
Community reportsimple pick, roughly half the time
10 demos
MolmoAct2practical threshold in its own docs
< 200 demos
GR00T new_embodimentofficial guidance, in steps not demos
unstatedgives steps, not demos
Four sources, four incompatible answers, no way to predict which applies to your machine. The first thing AVIOS will publish is this curve, measured honestly on one arm, one task at a time — including the runs that failed.
Built in the open
One SO-101, one task at a time, every footgun written down with what it cost in hours. If you own a cheap arm and cannot get a policy running on it, that is exactly the problem being worked on.
Cite this page
AVIOS: the install step for robot policies
AVIOS · October 2026 · https://avios.ai/
Concept page. Model facts checked October 2026. Nothing ships yet.
APA
AVIOS. (2026). AVIOS: the install step for robot policies. https://avios.ai/
BibTeX
@misc{avios2026,
title = {AVIOS: the install step for robot policies},
author = {{AVIOS}},
year = {2026},
url = {https://avios.ai/},
note = {Concept page. Model facts checked October 2026}
}
HTML
<a href="https://avios.ai/">AVIOS: the install step for robot policies</a>