Concepts

The vocabulary Euler uses, and what each term commits the platform to.

Episode

An episode is the canonical unit of robot-learning data in Euler. Raw logs become episodes with aligned streams, event labels, readiness evidence, annotation layers and exportable lineage. Everything else in the platform is a transformation of, or a judgement about, an episode.

Use-Case Workspace

Every workspace declares a use case: the problem it exists to solve, such as robotic manipulation, autonomous driving, drones and aerial, vision datasets, spatial and world models, or humanoids and mobile robots.

The use case themes the whole platform: the language screens use (images against episodes against flights against drives), which target profiles apply and which one is recommended, which source and export formats fit (every supported format stays visible, unfit ones are disabled), which readiness components get emphasis, and the default annotation layer modes a project inherits.

The catalog is served by GET /v1/use-cases. A use case is mandatory at workspace creation and can be set on pre-existing workspaces via PUT /v1/workspaces/{id}/use-case.

Dataset Kind

Where the use case describes the customer’s problem, the dataset kind describes the data itself: teleop, egocentric, drive-log, aerial-flight, state-only or vision-asset.

The kind decides which measurement contracts exist at all. An egocentric human clip has no commanded trajectory or gripper channel to score, a drive log never settles the way a demonstration does, and a state-only log has no cameras to drop. Readiness applicability, annotation layer applicability (hand pose runs on egocentric capture, not on drive logs), and the kind chip on every episode all read from one registry.

Derivation is automatic from episode content and resolved fresh on every response, so it can never go stale. Genuinely ambiguous projects can pin an override via PUT /v1/projects/{id}/data-kind.

Annotation Layer

A layer is one class of annotation output. Euler catalogs twelve, from scene captions to 3D representation, and eleven of them are servable today. Each is a catalog row recording what it produces, which models back it, which kinds can carry it, how it is QA’d, which export formats carry it, and the knobs a team can set.

A layer whose lane is not deployed stays in the catalog carrying that fact, rather than being dropped from the list. It is pinned off and cannot be queued. The 3D representation layer is in that state today.

The catalog is the single source of truth: the settings screen renders it, the run worker and the job queue consult resolved modes before spending, and the coverage report reports against the same rows. See annotation layers.

Layer Mode

Each layer is in_run (executes inside the pipeline run), on_demand (reachable afterwards through its job endpoint) or off (never spends). The effective mode resolves project settings over workspace settings over a computed use-case default. See pipeline settings.

Annotation Job

Heavy layers run through an asynchronous queue rather than a synchronous request, because the platform’s edge proxy caps a request at 300 seconds and the GPU lanes exceed it. The request returns 202 with a job record; the client polls or subscribes to an SSE stream. See annotation jobs.

Annotation Review

A separate pass that grades the annotations rather than producing them. It looks at sampled frames alongside what each layer produced and records a verdict per layer, with evidence, suggested improvements, and an explicit not-assessable where nothing could be judged. It runs as a job and has no pipeline settings tile. See automated review pass.

Readiness Components

Readiness is a weighted blend of twelve deterministic checks, renormalised over the ones that apply to the data at hand: timeline integrity, kinematic sanity over the action values, cross-signal synchronization for multi-rate recordings, length, gripper and language checks, camera coverage, image integrity for still-image assets, and annotation coverage.

Each component below its threshold raises a named flag that travels with the receipt. A project can disable a check, nudge a threshold within a bounded range, or override a weight. See readiness.

Readiness Layers

The same measurements grouped into three questions with three different owners: integrity (was it captured correctly), semantic quality (do the labels describe what happened) and utility (will this help the model you are training). The grouping recomputes nothing, so it can never disagree with the headline score.

The utility layer deliberately carries no single number, because blending its signals would need weights nobody has calibrated against a measured training outcome. Anything unmeasured reads as not assessable with the reason. See data utility.

Readiness Receipt

A Readiness Receipt is a verifiable summary of why an episode was accepted, rejected, or sent to review. It carries the quality score, decision, evidence, source hashes, and a stable content hash.

Euler Certified Report

An Euler Certified Report answers a different question from a dataset export: whether the current source evidence is suitable for a named training target under a pinned policy. It never rewrites or copies training media.

Each metric exposes the observed value, formula, inputs, provenance, confidence, applicability, ideal boundary and review boundary. The aggregate verdict is certified, review_required, not_ready or not_assessable; missing modalities reduce evidence coverage instead of silently becoming zero or pass.

Certification policies are immutable and content-hashed. A reviewer can adjust target-dependent defaults, but saving creates a new version and retains prior reports. See the data assessment guide.

Target Profile

A target profile is the consumption contract of the downstream training consumer: VLA, behaviour cloning, diffusion policy, world model, SSL encoder or detector. It is declared at project creation, drives which pipeline steps run (a detector target runs grounded object labeling), selects the export format (LeRobot v3, WebDataset shards, COCO detection, image folder, zarr replay buffer), and gates “ready for this target” validation, including an action-space diagnosis that recommends the matching policy class.

Model Registry

Every model-backed surface resolves its model through one registry: a default per surface, then an optional per-use-case override, then a tenant choice, then an environment override. Each row carries the benchmark evidence that justified it, so there are no unmeasured swaps, and the reason a higher-scoring alternative was rejected when it was rejected on license grounds.

Perception surfaces run self-hosted models only. No external label API is used, and customer frames never leave the platform boundary. Read the resolved mapping with GET /v1/models.

Search Index

The search index runs typed filters before vector ranking. That keeps customer queries auditable: task, outcome, object, failure mode, robot, review decision, tags and score thresholds stay visible even when vector similarity contributes ranking evidence.

Saved Slice

A saved slice freezes a search result into a reproducible definition. Saved slices preserve membership, the exact query, the coverage, the policy and the search-response hash. A slice can later be promoted into a locked eval set.

Curation Plan

A curation plan is an ordered sequence of strategies executed against a project’s episodes, producing a subset plus a report that attributes every dropped episode to the stage that dropped it. Strategies are registry rows, not code branches. See curation.

Euler Copilot

The Copilot is the agentic control plane: a tool-calling loop over a typed registry of platform capabilities, with role-gated tools, confirmation on destructive or spend-incurring actions, a bounded step limit, budget guards and a structured transcript. Customer frames never reach the chat provider. See Euler Copilot.

Closed Loop

Model feedback closes the loop: imported failure records are embedded and scored for novelty against the project’s stored embedding baseline (the training distribution). Failures near training data become an auto-slice of the nearest training gaps, curation re-runs with those episodes boosted, and a staged re-export waits for a one-click human seal.

MCP Surface

The Euler MCP server exposes agent-native tools for search, slice creation, guarded export planning, receipt verification and feedback import. Export actions that involve private data or large episode counts require human approval. See the MCP quickstart.

Roles and Access

Identities carry one role on an ordered ladder: viewer < reviewer < engineer < admin. The API enforces the ladder on every endpoint: ingest and pipeline runs need engineer or above, review decisions and dense-text edits need reviewer or above, and tenant administration (users, tokens, budgets) needs admin.

The app reflects the same ladder. A viewer gets a read-only product with the ingest surface and the trust section hidden. Viewer access is how customers inspect their own deliverables (episodes, receipts, overlays and exports) without any risk to sources or tenant configuration.