<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Notes on Robotics and AI on Research</title><link>https://research.revlev.org/</link><description>Recent content in Notes on Robotics and AI on Research</description><generator>Hugo</generator><language>en-US</language><lastBuildDate>Tue, 06 Oct 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://research.revlev.org/index.xml" rel="self" type="application/rss+xml"/><item><title>How AI read Napoleon's 1809 cipher letter — and how we checked it</title><link>https://research.revlev.org/posts/how-ai-read-napoleons-1809-cipher/</link><pubDate>Sun, 04 Oct 2026 00:00:00 +0000</pubDate><guid>https://research.revlev.org/posts/how-ai-read-napoleons-1809-cipher/</guid><description>In September 2026, GPT-6 Astra read a 217-year-old Napoleonic cipher from one scan. We downloaded the solution package and ran it offline: the key reproduces the literal text byte for byte, and 97% of the letters reach the final reading unchanged.</description></item><item><title>When cutting power isn't safe: a ROS 2 safety supervisor for ground robots and UAVs</title><link>https://research.revlev.org/posts/when-cutting-power-isnt-safe/</link><pubDate>Wed, 30 Sep 2026 00:00:00 +0000</pubDate><guid>https://research.revlev.org/posts/when-cutting-power-isnt-safe/</guid><description>Most robot safety designs quietly assume that removing power is a safe default. On a ground robot it is. On a drone it&amp;rsquo;s a free fall. This post walks through a ROS 2 safety supervisor that keeps what must stop separate from how to stop it — from heartbeat silence and criticality classes to a hardware fallback that takes over the propulsion signal.</description></item><item><title>Cheap model first, stronger model when it matters: escalating on validation failure in Pydantic AI</title><link>https://research.revlev.org/posts/escalating-on-validation-failure/</link><pubDate>Sun, 27 Sep 2026 00:00:00 +0000</pubDate><guid>https://research.revlev.org/posts/escalating-on-validation-failure/</guid><description>Most extraction requests don&amp;rsquo;t need a frontier model: a cheap one gets them right, and validators catch the rest. This article shows how to retry the same model with the error, then escalate to a stronger model that sees what went wrong, using a small Pydantic AI capability, pydantic-ai-escalation. It covers why the framework&amp;rsquo;s built-in fallback doesn&amp;rsquo;t fit, why counting failures correctly is harder than it looks, and two production traps the tests caught.</description></item></channel></rss>