Florian Krebs

At a glance

The sixteen-year line

The work shepard inherits is not a fresh start. It is the latest in a sixteen-year continuous line of work on the data-substrate problem of full-scale composite manufacturing, all conducted at the same site — the DLR Center for Lightweight Production Technology (ZLP) in Augsburg.

The lineage runs through three substrates:

• 2010 — MFZ Steuerungskonzept: the control-concept document for the centre’s flagship Multi-Function Cell. The first recovered primary source establishing the author already working on the data side of the centre’s flagship robot cell, one year after joining DLR.

• 2017–2020 — CUBE / iDMS: a Krebs-architected research-data prototype used in the IPRO project. Neo4j + MongoDB + InfluxDB substrate stack, Project / Experiment / Step entity model, PROV provenance. The direct architectural ancestor of shepard.

• 2021–present — shepard: the open-source substrate. Lead author on the peer-reviewed architecture paper of record (Krebs et al., DLRK 2021); named maintainer of the noheton fork the thesis work is grounded in.

The trajectory is unusually coherent: predecessor systems (PRAESTO commercial, KIBID vendor-built, iDMS prototype) failed by an enumerable set of mechanisms his earlier work documents, and the substrate now shipped — plugin-extensible, ontology-driven, AI-collaboratively developed, federation-ready — is the structured response to each named failure.

For the full predecessor history, see the upstream upgrade path.

Industrial pedigree

Before DLR, April 2008 to September 2009: a 17-month industrial year at KUKA Roboter GmbH, Augsburg — the vendor whose KR-series robots later anchor every MFZ, FPC, and T-AFP cell at ZLP. The clean industrial-pedigree → research-substrate arc passes through here. The KUKA period also intersects with the European research consortium that would later produce the Multifunctional Fuselage Demonstrator (MFFD, JEC World Innovation Award 2025 — Aerospace Parts).

Education: Diplom-Informatiker (Applied Computer Science), Universität Augsburg, 2002–2008. Diplom overall grade 1.38 (sehr gut) — Hauptstudium weighted average 1.51, Diplomarbeit graded 1.0, Philosophy minor 1.94. The earliest peer-reviewed publication on ORCID — Relational Cognitive Structures for Intelligent Agent and Robot Control (IEEE SMC 2008) — predates DLR employment by ~1 year and signals the ontology-first design posture that surfaces throughout the later work.

The specialisation tracks (Vertiefungsbereiche) covered Multimedia & AI (top-graded KI-Praktikum 1.0, Multimediapraktikum 1.0, Spieleprogrammierung 1.0), Systemnahe Informatik (Mikrorechnertechnik I 1.7 + II 1.3, Praktikum Prozessorbau 1.3, UbiCom Praktikum 1.0), and Theoretische Informatik (Algebraische Methoden / Topos theory, Verteilte Algorithmen, Algebraische Beschreibung paralleler Prozesse). Coursework also included Advanced Cognitive Robotics and Intelligente autonome Roboter — both graded 1.0, the natural precursor to the 2008 IEEE SMC robotics paper and the subsequent KUKA industrial year. Bio-inspired & Grid Computing (graded 2.3) and Organic Computing sit in the same emergent-systems line.

The Nebenfach (minor) was Philosophy — graded coursework across Logik / Sprachphilosophie, Erkenntnistheorie / Kognitionswissenschaften, Philosophie des Geistes, and Ethik / Praktische Philosophie. The specific seminar palette, with lecturers and grades, recovered from primary-source archive:

• “Natur, Ordnung, Geist” — Vorlesung, Prof. Dr. phil. Severin Müller (Uni Augsburg philosophy faculty).
• “Philosophie des Geistes — Grundlagen der Erkenntnistheorie, Künstlichen Intelligenz, Wissenschaftstheorie” — Vorlesung, Klaus Mainzer.
• “Das Unendliche. Von Logik, Mathematik, Gott, Welt und Computern” — Seminar, Mainzer, graded 2.0.
• “Naturphilosophie: ausgewählte Themen, insbes. Philosophie der Biologie” — Seminar, Justus Leiber, graded 2.0.
• “Das Ende des Menschen? Philosophie, Anthropologie und TechnoScience des Posthumanismus” — Blockveranstaltung, Leiber, graded 1.0.
• “Verantwortung. Zur aktuellen Debatte über einen neuen Grundbegriff der Ethik” — Seminar, Schröer, graded 2.7.

The intellectual genealogy of the present fork’s posture — ontology-first, AI-collaboratively developed, provenance-disciplined — has its roots here. The combination is unusually specific: a Mainzer lineage on cognition + complex-systems epistemology, a Leiber lineage on philosophy of biology and posthumanism (with the Das Ende des Menschen? seminar on posthuman TechnoScience graded 1.0, the load-bearing top mark of the minor), and a Schröer lineage on Hans-Jonas-style ethics-of-responsibility — three lecturers each anchoring one of the three argumentative spines the present AI-collaborative thesis method operates on.

The Diplomarbeit project (2007–2008) was PARIA — “Entwicklung eines Systems zur Implementierung von Layered Module Architekturen”, supervised by Dr. Martin Eric Müller at the Lehrstuhl für Multimedia-Konzepte und ihre Anwendungen, Universität Augsburg (today at H-BRS, adjunct at Uni Augsburg). The Diplom was not an isolated project but the culmination of three years of lab work at that chair: Krebs first met Müller in the Künstliche Intelligenz lecture in SoSe 2005, joined the chair’s multi-agent robotics framework as the developer responsible for the server components, and independently built a fully transparent interactive simulation environment so the robot-control intelligence could be developed and tested without hardware in the loop. The lab role was paid as a Hilfskraft in WS 2007/2008; Müller had earlier nominated Krebs for the Studienstiftung des deutschen Volkes in 2006. Technically: a modular reactive-agent runtime grounded in Rodney Brooks’s Subsumption Architecture (1986), implemented on a Lego Mindstorms RCX substrate with Java-API design discipline (Javadoc-documented exceptions: ConnectionException, RobotException, TimeoutException, RCXException, RallyRobot), plus a JabRef-managed bibliography running from Braitenberg’s Vehicles through Agre & Chapman, Brooks himself, and the 1990s connectivist-agent line. The architectural posture — layered, modular, reactive over deliberative, the world as its own model — is recognisably the same posture the shepard plugin-SPI substrate now embodies sixteen years later. The robot-control + provenance discipline that runs through the later DLR work has its first production-shape here.

Things to explore deeper

Three threads surfaced by the UNAS-archive deep-dive that warrant further primary-source work in subsequent dossier rounds:

1. Two Müllers, not one. An earlier dossier round misidentified the Natur, Ordnung, Geist lecturer as the Diplomarbeit advisor and built a “single supervisor bridging philosophy and engineering” narrative on it. Primary-source correction: those are two distinct people — Prof. Dr. phil. Severin Müller on the philosophy side, Dr. Martin Eric Müller (CS, Lehrstuhl Multimedia-Konzepte 2005-2008) on the advisor side. The dual training in philosophy and Brooks-school engineering is real, but it came via two separate teachers, not one institutional joint.

2. Brooks → AI-collaborative method, the load-bearing arc. The Diplomarbeit argues reactive agent architectures outperform planning systems in unknown, fast-changing environments. The present thesis-method posture — AI-as-collaborator-rather-than- oracle, durable per-turn provenance, the conversation is the data — has a structurally identical shape: prefer reactive exploration with a faithful trace over up-front deliberation. The line from Brooks 1986 through Krebs 2008 PARIA through Krebs 2026 f(ai)²r deserves an explicit chapter-level treatment.

3. The KUKA 17-month window (April 2008 – September 2009). Still the only Krebs career segment with no primary-source content in this profile. What product family / customer engagement was the industrial year on? Tracked as RL-UNAS-KUKA for the next dossier round.

Selected publications

A small selection from the ORCID work list; full bibliography on /shepard/bibliography.

• Krebs et al. (2021) — Systematische Erfassung, Verwaltung und Nutzung von Daten aus Experimenten — DLRK 2021. The load-bearing peer-reviewed publication of record for shepard’s architecture; lead-authored jointly with the named software citation authors (Willmeroth, Haase, Kaufmann, Glück, Deden, Brandt, Mayer).

• Audet, Fortin, Côté, Vistein, Brandt, Krebs et al. (2024) — Iterative Robot Calibration with Accuracy Saturation Termination — ICRA 2024 (IEEE). ICAST algorithm; in-situ elastic calibration of the DLR ZLP T-AFP cell; joint DLR–NRC publication.

• Krebs (2026) — Generative KI: Technologie, Ethik und Zukunft — Wie LLMs (unser) Arbeiten und Denken (verändern) — 17-slide author-led tutorial with explicit AI-Disclosure slide; the methodological deck for the AI-collaborative research method.

• Schönheits & Krebs (2018) — Embedding AR in Industrial HRI Applications — VAM-HRI 2018 (ACM). Industrial HRI on KUKA KR120 Quantec extra HA + HoloLens + UWB-RTLS + OPC-UA.

• Krebs & Jung (2013) — Hightech auf vier Säulen — Glanzstück im DLR Augsburg — DLR magazin #140. Public-facing institutional exposition of the MFZ at commissioning.

Method

shepard’s development is explicitly AI-collaborative under the author’s publicly defended position. Conversations with the AI assistant are kept as durable artefacts (a working memory of ~500 files persists across sessions); design decisions cite the prompt that produced them; provenance for every AI-generated contribution is captured through the f(ai)²r vocabulary the substrate itself operationalises. The thesis this work serves is reflexively a worked example of the very pattern its substrate is built to capture.

Beyond research — dive log

A long-standing diving practice runs alongside the research line: 243 logged dives across seven countries, 2009–2015, tracked in Subsurface. The map below shows country-level aggregation — circle radius scales with dive count. Site-level coordinates and individual dive records are intentionally not surfaced; only the country centroid and the per-country tally.

The distribution is itself characterful: 108 of 243 dives (44%) in the Red Sea / Ägypten — a return-site pattern rather than a tourist-trail spread — followed by a Southeast-Asia cluster (Malaysia, Philippinen, Indonesien, jointly 98 dives) covering Sipadan, Lankayan, Komodo, Anilao, and Apo Reef. The seven-country footprint includes one home-country line (Deutschland, 10 dives, mostly Alpine lake / Bodensee). The dataset stops in late 2015 — a marker, in passing, of the work-life boundary the DLR period imposed.

Get in touch

• ORCID — https://orcid.org/0000-0001-6033-801X (canonical)
• LinkedIn — https://www.linkedin.com/in/florian-krebs/
• Bluesky — https://bsky.app/profile/nucli.de
• GitHub — https://github.com/noheton
• Email — florian.krebs@dlr.de

For the detailed scholarly profile (timeline, co-authorship network, methodological signature, current projects), see aidocs/strategy/104-author-research-profile.md in the repository.