Mosaico (@MosaicoLabs) header
Mosaico (@MosaicoLabs) profile picture

Databricks for Physical AI

Joined February 2026
Photo by Mosaico (@MosaicoLabs)Photo by Mosaico (@MosaicoLabs)Photo by Mosaico (@MosaicoLabs)Photo by Mosaico (@MosaicoLabs)Photo by Mosaico (@MosaicoLabs)
Applications are open for Physical Education Vol. 1. 🦾 A physical AI hackathon on real SO-101 arms: teach a robot a task with your own data, then watch it run on its own. 📍 Albert School, Milan · Nov 27–29 🎟️ Free, limited seats 👉 luma.com/mwu3zj7i
164
🚨 BREAKING:  Italian startup @MosaicoLabs has raised €7.5M in seed funding to fix one of the least glamorous problems in physical AI: robotics data. The round was led by United Ventures, with OpenOceanand existing investor Connect Ventures. Total funding is now €8.5M, following a €1M pre-seed in February. Everyone talks about models and robots. Far fewer talk about what happens to the data in between. A robot fleet produces huge volumes of sensor data. Before any of it can train a model, someone has to organise it, label it and find the right examples. Today that is mostly manual work. Teams can spend hours digging through datasets to find a specific scenario. Mosaico is an open-source platform that automates that work. → Automatically organises and processes data from robotic sensors → Search in natural language, so results come back in seconds rather than hours → Example: ask for every instance of a vehicle stopping at a pedestrian crossing, instead of reviewing footage by hand → Plugs into existing systems, with no dedicated data infrastructure to build The founders are three former colleagues from Ambarella, where they spent 15 years building the full autonomous vehicle pipeline. CEO Gabriele Lini says they understood the “manual data plumbing problem” first-hand. That is the right background for this product. The open-source community is the traction to watch. It has grown to 1,200+ members since December 2025, including engineers and researchers from Google DeepMind, Amazon and ETH Zürich. Mosaico is also honest about the market. Many potential customers are not yet ready to buy specialised data software at scale, so the focus for now is developer adoption and becoming a standard for robotics data management. Scaling laws say more data means better robots. That only works if you can find, filter and use the data you already have. The picks-and-shovels layer of physical AI is getting built, and some of it is coming out of Parma. 🇮🇹 Check the repo out here: github.com/mosaico-labs/mosa… ~~  ♻️ Join the weekly robotics newsletter, and never miss any news → ziegler.substack.com
13
4
5
81
7,880
Software learned this slowly and expensively. For years, debugging a production issue meant grepping through raw logs, reconstructing what happened from flat output, and hoping you found the right line. It worked until the systems got complex enough that it didn't. Then structured logging happened, error classification happened, queryable observability happened, and nobody went back. Robotics is still in the grep phase. ~1.5 TB per robot per day. Bag files. No structure, no classification, no index. When a fault happens => open files manually, align timestamps, write scripts, spend the afternoon. The data exists. It just can't answer questions. The pattern that replaced grep in software is emerging in robotics. Classify faults at the source. Capture structured context automatically. Ingest as a queryable record, not a file. The Selfpatch team ran this last week on a real robot. LIDAR fault, classified at trigger, 25 seconds of context auto-captured, ingested into @MosaicoLabs with one call. Three diagnostic questions in thirty seconds instead of an afternoon. 414 MB per fault vs 1.5 TB/day. ~3,500x less storage. The transition in software wasn't announced. Teams just started doing it because the old way stopped working. The same thing is starting to happen in robotics.
26
23
76
2,607
🚨Mosaico v0.3.0 is now available! This release introduces updates to the platform stability and security framework. Key Highlights in v0.3.0: (thread)
1
93
4) Main Branch Migration: we've moved our primary development to the main branch to streamline community contributions.
1
11
The official Mosaico documentation is now live. We’re building the industrial-grade data engine for Physical AI—moving beyond messy logs toward structured, searchable, and AI-ready sensor data. 🚀 📖 Docs: docs.mosaico.dev
1
1
74
Why does this matter? Most robotics teams spend 50% of their time fighting data pipelines. Mosaico standardizes ingestion, storage, and retrieval so you can focus on behavior, not infrastructure.
1
1
52
We are keeping the core engine open source (AGPL-3.0) because the future of robotics depends on shared standards. Go ahead, break things, build things, and let us know what you think! 🎸 ⭐️ GitHub: github.com/mosaico-labs/mosa…
1
15
Mosaico is officially Open Source! Find the code on GitHub and give us a ⭐️ to support the project. github.com/mosaico-labs/mosa…
1
105
The Open-Source infrastructure for Physical AI Data Analysis Happy to announce the strategic partnership between @MosaicoLabs and PlotJuggler, the open-source time-series visualization tool created by @facontidavide. Check out the Mosaico Repo: github.com/mosaico-labs/mosa…
1
2
55