Quickstart

This is the shortest practical path through the platform. The goal is not to force full migration on day one. The goal is to prove value with one real workflow while preserving the structure needed to expand into a full lifecycle system later.

Step 1 — Connect A Data Source

Start with the least disruptive entry point that matches how your team already collects data.

robot-side agent for continuous operational collection

direct cloud import for historical or edge-synced data

API, SDK, or CLI for developer-controlled ingestion

Platform surfaces involved

Robot Data Ingestion
Platform API
Identity, Access & Governance

What matters here

Capture robot identity, operator context, site or environment metadata, and source lineage at intake. That context becomes important much later when training, deployment, or incident analysis needs to trace back to the source.

Step 2 — Process And Review The Data

Once data lands in the platform, process it into canonical episodes and inspect it in synchronized replay.

validate streams and timelines

inspect sensor completeness and alignment

review failures, anomalies, and obvious low-value sessions

Platform surfaces involved

Data Processing Pipeline
Data Explorer
Workflow Engine

What matters here

Processing should separate upload reliability from data correctness. Review should happen on the same structured entities that later labeling and curation will use.

Step 3 — Label What Matters

Apply the annotation schema needed for the task and the release process.

episode-level success and quality labels

temporal phase and failure window labels

frame or spatial labels when the task needs them

structured review and approval of labeling quality

Platform surfaces involved

Annotation & Labeling
Data Explorer
Identity, Access & Governance

What matters here

Labels should remain attached to source data, schema version, and review state so they stay usable when data is reprocessed or promoted into new dataset versions.

Step 4 — Curate For Coverage, Not Just Volume

Use quality scoring, retrieval, and policy-based selection to choose what should influence training.

remove obvious garbage and low-signal duplication

mine failures and intervention-heavy episodes

balance by task, site, operator, or hardware revision

build reproducible curation ruleset instead of one-off selection

Platform surfaces involved

Data Curation Engine
Dataset Finalizer
Data Explorer

Step 5 — Freeze A Dataset Snapshot

Finalize the selected data into an immutable dataset object with explicit split and manifest policy.

freeze exact membership

attach annotation and curation lineage

validate consistency and completeness

export manifests for training and evaluation systems

Platform surfaces involved

Dataset Finalizer
Unified Data Model
Workflow Engine

What matters here

The platform should hand training a governed dataset object, not an ambiguous folder path.

Step 6 — Train And Compare

Launch training from the finalized snapshot and compare candidates with explicit experiment structure.

run local, cloud, or hybrid compute

track checkpoints, config, and runtime metadata

compare runs inside experiments

promote only meaningful candidates into the registry

Platform surfaces involved

Training Orchestrator
Experiment Tracker
Model Registry

Step 7 — Evaluate And Release

Before rollout, test the candidate against replay packs, benchmark suites, and promotion policy.

compare against baseline

apply thresholds for success, latency, and intervention budget

require approval where risk warrants it

Platform surfaces involved

Evaluation & Release
Model Registry
Workflow Engine

Step 8 — Deploy, Monitor, And Close The Loop

Package the approved model, stage rollout, monitor the cohort, and feed field evidence back into the next development cycle.

build artifact

canary rollout

monitor telemetry and intervention rate

create incidents or maintenance cases when needed

promote valuable field evidence into new data and evaluation coverage

Platform surfaces involved

Artifact Builder
Deployment Manager
Update Manager
Fleet Management
Telemetry & Monitoring
Maintenance System
Human-in-the-Loop Operations

Early Decisions Worth Making Carefully

Primary ingestion path

Choose the entry point that matches today’s operations, not an imaginary end state.

Annotation schema

Define the task ontology before scaling collection so labeling stays consistent.

Curation policy

Decide what “good data” means in a reproducible way.

Release criteria

Define benchmark and promotion rules early so teams are not inventing standards after the candidate already exists.

Residency and compliance boundary

Decide early whether the deployment needs region-bound storage, cross-border transfer controls, sovereign hosting, or restricted teleoperation and support access, because those decisions affect ingestion, API access, rollout topology, and operations workflows.

What A Good Pilot Looks Like

A strong first implementation proves all of the following:

data enters with real provenance

teams can review and label it in one place

datasets become immutable, explicit objects

training is reproducible

release decisions use structured evidence

rollout and field feedback remain connected to the same lineage chain

That pilot is the best foundation for expanding the rest of the lifecycle into the platform.