{"slug": "introducing-foxglove-datasets", "title": "Introducing Foxglove Datasets", "summary": "Foxglove launched Foxglove Datasets, a feature that turns time ranges in robot recordings into versioned training data without copying the underlying sensor data, available on Foxglove-hosted, Bring Your Own Storage (BYOS), and self-managed deployments on every plan. Datasets builds on Foxglove's existing search, events, and Automations tools to convert data moments into episodes, organize labeling information, and provide statistics, embedding maps, versioning, and lineage across the ML data lifecycle into PyTorch or Ray. Foxglove said the feature addresses ML teams that were curating data in one tool, versioning in tools like DVC, and writing one-off scripts, making it hard to trace which recordings, time ranges, and labels went into a given dataset or model.", "body_md": "# Introducing Foxglove Datasets\n\nTurn time ranges in your recordings into versioned training data, without copying the sensor data.\n\nToday we are introducing Foxglove Datasets to help turn your robotics data into high-performance models. We built Datasets because a lot of the ML teams we work with were already using Foxglove to find the moments that matter in their data, and then leaving Foxglove to turn those moments into training data. We want our data platform to support the full ML data lifecycle: going from raw robot data, to episode curation and labeling, to dataset curation and evaluation, to ML-ready training data that goes straight into PyTorch or Ray. All of this is backed by strict versioning and lineage from one end of the data lifecycle to the other.\n\nIn this post, we’ll cover the problems we kept hearing about from ML teams, how we designed Datasets to solve them, and how you can get started today. The Datasets feature is available on Foxglove-hosted, Bring Your Own Storage (BYOS), and self-managed deployments, on every plan.\n\n## The questions every robot-learning team asks\n\nIn our conversations with customers, we have found that teams building models on robot data keep running into the same questions:\n\n- How do we structure, store, and organize the data we need for training?\n- How do we store it cheaply, without paying for copies?\n- How do we let several people curate data without stepping on each other?\n- How do we ensure data transfer is not a training bottleneck?\n- How do we explore the lineage of data that a model was trained on?\n- How do we trace model misbehaviors back to our data?\n\nOur customers were answering these in a variety of ways: curating data in one tool, versioning in tools like DVC, one-off scripts to pull whatever is needed at a specific point in time, or building and maintaining their own internal tooling. Once the data is spread across work streams like this, it is a challenge to track which recordings, time ranges, and labels went into a given dataset and model, and how that data was collected in the first place.\n\nWhen a model regresses or needs improvement, the first thing someone has to do is work out what data it was trained on and how that data was collected, in order to gather additional data and refine existing data to improve the model. Without a long-standing history, that can be painful, which is why Foxglove decided to tackle this problem.\n\n## How Datasets fits into the Foxglove data platform\n\nDatasets builds on the rest of the Foxglove data platform. Your recordings stay where they are. You find useful data moments with [search](https://docs.foxglove.dev/docs/data/search) and [events](https://docs.foxglove.dev/docs/data/events), and now [Automations](https://foxglove.dev/blog/how-we-built-foxglove-automations-and-what-comes-next) for automatic processing as new data comes in. Datasets turns those moments into episodes, help organize labeling information, provide statistics and embedding maps to understand the distribution of your data, and provide versioning and lineage for all training work.\n\n### Modeling episodes without duplicating data\n\nBuilding a training set usually means moving data between tools. Recordings get pulled out of storage, converted into a training format, and copied to wherever training runs, and someone has to keep track of which copy went into which dataset. That takes engineering time, and there is cost to store those extra copies. We wanted episodes to be easy to make and cheap to store at scale.\n\nWe knew performance and cost would be the two things that mattered most here, and that informed how we modeled episodes. An episode represents one example, from a single image frame for object detection to a sequence of multi-modal data demonstrating an action. Episodes specify a time range and reference recordings, building off of our existing streaming infrastructure, which allows teams to quickly iterate on episodes without duplicating the underlying data. Many episodes can overlap on the same two-hour recording, and the sensor data is still only stored once.\n\nSearch allows for querying deep into your data, and the episode tooling allows for visualization and fine-grained control over trimming, so you can go from a search result to a trimmed episode quickly. Custom properties can be tracked alongside episodes, supporting data such as labels and tags.\n\n### Versioning\n\nAs you train and evaluate models, it’s important to be able to understand the data that went into them. When we asked teams how they tracked this, the answers varied: some hashed their files, some bumped a version number every time labels or metadata changed, and some didn’t version at all.\n\nThe main decision we had to make was when a version gets created. We thought about versioning every change automatically, the way a shared doc does, but the teams we spoke with cared about controlling the curated set of versions that connected to the new iterations of their models. So edits go into a shared draft and you commit a version when you’re ready, which for most teams is right before a training run. Each committed version is immutable and records who committed it and when, as well as statistics that explain the distribution of data.\n\nEpisodes also change after a dataset has been used for training, since people trim them, fix labels, and add new types of labels to the same episode. To handle that, all changes to datasets and episodes are versioned, and a committed dataset version keeps pointing at the episode versions it was committed with. That is what allows you to return to the exact state of a dataset when a specific model was trained, even if the episodes in it have been edited since.\n\nWe also believe you should look at a dataset as a whole before you spend GPU time on it. Problems like a dead camera, a lopsided mix of tasks, or too many near-identical episodes are easy to see across a dataset and expensive to find after a training run. So from a dataset you can explore the distribution of its episodes. Histograms allow for easy identification of outliers, or gaps where more data may be required. Our new embedding view makes it easy to see clusters across visual data, highlighting where there are too many similar samples or not enough to have a representative dataset.\n\n### Training\n\nWhen it comes to training, most teams use an ML framework like PyTorch or Ray, so we started with data loaders for these two packages. This enables engineers to easily interface with Foxglove datasets via Python notebooks and their existing training code.\n\nA key for efficient training is to perform as little transfer and manipulation of data as possible while training runs. The data loaders help with filtering datasets by properties, limiting data to select topics, and selecting splits ahead of training. They pull only the necessary data directly from object storage, decompress video, and align frames.\n\n## An example: teaching a bimanual robot to pick fruit\n\nWe recently bought a [Trossen Mobile AI](https://www.trossenrobotics.com/mobile-ai), a bimanual robot with two leader arms and two follower arms, and set out to train a policy to pick up and sort fruit. For our first dataset version, we marked each attempt with an event and turned the events into episodes. After that, we had the robot publish a topic that is true whenever an attempt is in progress, and used search to turn those ranges into episodes.\n\nOur first policy couldn’t finish the task consistently. Looking through the dataset, we saw that all of our demonstrations were recorded on a table at a different height than the one we tested on, so the arms reached the fruit but couldn’t find the right height to grasp it. We recorded new attempts on the new table, committed them as a new dataset version, and retrained on it. The first version stayed as it was, so we could compare the two.\n\nMostafa Elshenawy walks through the whole thing, including the code, in [Training a Bimanual Policy on Foxglove Datasets](https://foxglove.dev/blog/training-a-bimanual-policy-on-foxglove-datasets).\n\n## Getting started\n\nDatasets works with data that’s already in Foxglove, so there’s nothing to migrate to get started. Creating datasets and dataset versions is free. Episodes count toward your monthly usage one time, when they’re created, and every plan includes a monthly allowance. The [pricing docs](https://docs.foxglove.dev/docs/pricing#episodes) have the details.\n\nHere’s what we’d suggest:\n\n- Start small. Pick the failure you most want a model to stop making, and find examples from events, search results, or recordings.\n- Commit a version when you train, and link that version with your checkpoint. If a model gets worse, you can open the version it was trained on and see exactly what changed.\n- Click **Use for training** to generate code for PyTorch or Ray, or**Copy prompt** to give the same details to your coding agent.\n- Use [Automations](https://foxglove.dev/blog/how-we-built-foxglove-automations-and-what-comes-next) to find candidate moments as data comes in and continuously add them to a draft dataset version.\n\nWe’d love to hear how Datasets works for you. The [technical walkthrough](https://foxglove.dev/blog/training-a-bimanual-policy-on-foxglove-datasets) and the [Datasets docs](https://docs.foxglove.dev/docs/data/datasets) are good places to start, and you can share feedback in our [community](https://foxglove.dev/chat).", "url": "https://wpnews.pro/news/introducing-foxglove-datasets", "canonical_source": "https://foxglove.dev/blog/introducing-foxglove-datasets", "published_at": "2026-11-17 18:00:00+00:00", "updated_at": "2026-11-17 19:58:34.946060+00:00", "lang": "en", "topics": ["robotics", "machine-learning", "ai-infrastructure", "mlops", "developer-tools"], "entities": ["Foxglove", "Foxglove Datasets", "PyTorch", "Ray", "DVC"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/introducing-foxglove-datasets", "markdown": "https://wpnews.pro/news/introducing-foxglove-datasets.md", "text": "https://wpnews.pro/news/introducing-foxglove-datasets.txt", "jsonld": "https://wpnews.pro/news/introducing-foxglove-datasets.jsonld"}}