Make it yours
The Trossen dataset was a stand-in for yours. Everything you walked through transfers directly, and you don't have to write it from a blank file.
Let the skills do the heavy lifting
Back in Foundations you installed Rerun's Agent Skills. They're how you point this course's patterns at your data: each one teaches your coding agent the exact API for a stage, so "convert my MCAP" or "export to LeRobot" produces Rerun's real patterns instead of guesses. Install all of them, or just the one you need:
npx skills add rerun-io/rerun # all of them
npx skills add rerun-io/rerun --skill rerun-mcap # just one| When you're… | Reach for |
|---|---|
| Deciding how your data maps onto Rerun (read this first) | rerun-data-model |
| Converting MCAP (incl. custom protobuf topics) | rerun-mcap |
| Deriving robot transforms from a URDF | rerun-urdf |
| Building a conversion or enrichment pipeline | rerun-chunk-processing |
| Exporting a curated slice to LeRobot | rerun-lerobot |
| Designing the dataset's default layout | rerun-blueprint |
Point it at your robot
| In the course | On your stack |
|---|---|
episode_*.mcap, opened natively or converted | Your existing logs: MCAP, ROS 2 topics, LeRobot datasets, or a custom loader for in-house formats |
Fixing calibrations and deriving /tf from URDF during conversion | The same patterns over your schemas, URDFs, and sidecar files |
| Registering recordings + a dataset blueprint | One dataset per robot configuration; one blueprint the whole team shares |
| Keyframe, blur, and metadata layers | Your real signals (tracking loss, reward-model scores, review verdicts, embeddings) attached as layers, never mutating raw data |
| Success-by-operator aggregate | Eval dashboards: success by model version, scene, or site, with results stored back as tables |
| Curated LeRobot export | Your training export, curated by query, re-runnable per experiment |
The companion code is the demo repo: clone it, run the local flow against the sample data, then start substituting your own MCAPs. Almost everything you saw is configurable, starting with the viewer layout itself: see Blueprints and Configure the viewer. The deeper query how-tos live in Query & transform, and the examples gallery covers many robot morphologies and sensors.
What a working data layer feels like
A checklist, honestly earned. You saw each of these this course, on real robot data:
- Anyone on the team can open any recording in seconds, and it arrives with a sensible layout.
- Recordings are searchable by the metadata that matters: robot, operator, task, model version, custom tags.
- Derived signals and annotations land after collection, as layers, without touching raw data.
- "Build a training set" means writing a query and a column mapping, not running a pipeline that copies data.
- An eval regression traces to specific recordings, and those recordings explain it.
When those are true, the experiment loop spins on data work instead of grinding on glue code. That gap, between teams that have this and teams that rebuild it quarterly, is the subject of The data layer tax for robot learning, which maps the whole problem walking backwards from evaluation to collection.
Keep going
- Point the Collect patterns at one real recording from your robot today. A single MCAP file is enough to start.
- The Discord is where "how do I log my weird sensor" questions get answered, often by someone who had the same sensor.
- And when your loop outgrows one machine, the same code runs against the whole fleet on Rerun Hub, the same data model you just learned, bigger iron.
You came in with a folder of files. You're leaving with a loop. Go spin it on something real.