Euler Docs
Robot logs in, training-ready datasets out
Euler is a data-readiness compiler for physical AI teams. It ingests raw multimodal robot logs, normalises them into episodes, scores every episode against deterministic readiness checks, annotates them across eleven servable layers, and emits versioned datasets with a per-episode Readiness Receipt that carries the evidence.
IngestMCAP, ROS-style dumps, Parquet, MP4, HDF5 and directory sources.
AnnotateEleven servable layers, each with its own QA receipt and export format.
MeasureTwelve readiness checks, gated by what your data can physically carry.
TrustReview evidence, save slices, export datasets, verify receipts.
Start here
15-minute quickstartCopy-paste curl, Python and TypeScript for the three core flows: register a source, plan a run, search and hand off a slice.ConceptsEpisodes, data kinds, readiness, receipts, target profiles, slices and the closed loop.Annotation layersWhat each layer produces, which data kinds carry it, how it is QA’d, and which one is not servable yet.Pipeline settingsSet each layer to in-run, on demand or off, configure it, apply a preset, and see the cost before you spend.ReadinessEvery check, what it means for training, and what failing it does to a policy.Data utilityReadiness split into integrity, semantic quality and utility, plus what has not been measured and why.Review workbenchOne recording at a time: every layer’s state, independent overlays, and per-recording runs.Euler CopilotThe agentic surface, what it can do, and the privacy boundary around your frames.
What Euler is
Euler is not a generic labeling dashboard. It is a compiler for training data: the same input produces the same output, and every judgement it makes is attributable to evidence you can inspect.
That shows up as four properties every customer-facing artifact has:
- Content-addressed. Sources, inventories, quality reports and receipts are immutable models with a
content_hash. Lineage is the record of which hashes produced which artifact. - Applicability-gated. A recording is never scored on a contract its class of data cannot have. An egocentric human clip has no commanded trajectory to complete; a drive log has no gripper events to be sane about. See data kinds.
- Reproducible. Saved slices preserve membership, the exact query, the coverage and the search-response hash, so an eval set does not drift under you.
- Evidence-first. Every model swap in the platform carries a benchmark reference, and every license-blocked alternative is recorded with the reason. See the model surfaces.
- Honest about gaps. Nothing unmeasured is reported as a zero or a pass. A check that could not apply, a metric with no reference behind it, and a claim with no measured run behind it all read as not assessable, with the reason and what would unlock them. See data utility.
Where to go next
- Running a first integration against the hosted API: Get started.
- Wiring an agent to Euler over MCP: MCP quickstart.
- Deciding whether a dataset is fit for a named training target: Data assessment.
- Selecting a training subset from what you have: Curation.
The hosted API is at https://api.euler.sudotank.com. The versioned contract is committed at openapi/euler.openapi.json.