Workflow Patterns
The Workflow Engine exists to make lifecycle execution explicit. These patterns are the ones robotics teams typically need first because they cut across multiple layers and usually break down fastest when handled through scripts, dashboards, and chat-based coordination.
1. Data To Training
This is the core development loop.
- ingest new sessions
- process and validate them
- review and annotate required labels
- curate for quality and coverage
- finalize a dataset snapshot
- launch training
Why this workflow matters
It formalizes the exact path from raw data to reproducible model input. Without it, the platform may hold all the right data but still rely on manual coordination for the most common lifecycle motion.
2. Candidate To Release
This workflow turns a promising checkpoint into a governed deployment decision.
- register candidate model
- attach evaluation pack
- run replay and benchmark checks
- compare against baseline
- request approval if policy requires it
- build artifact and mark it eligible for rollout
Why this workflow matters
It prevents release quality from depending on a mix of notebook screenshots, memory, and ad hoc Slack approvals.
3. Staged Rollout
This workflow bridges release confidence and production activation.
- select target fleet or cohort
- deploy canary
- observe health signals
- promote to wider cohort or pause
- rollback if policy thresholds fail
Why this workflow matters
A deployment is not one action. It is a monitored state transition with branching paths based on real field evidence.
4. Fleet Hotfix And Rollback
This is the fast path for operational recovery.
- detect incident or regression
- identify affected cohort
- evaluate rollback or corrective artifact
- trigger rollback or emergency deployment
- open follow-up incident and maintenance actions if needed
Why this workflow matters
Serious platforms need a governed emergency path, not only an ideal happy path.
5. Human-In-The-Loop Correction Loop
This workflow turns field interventions into better future models.
- detect intervention-heavy or anomalous sessions
- capture correction context
- review and annotate intervention phases
- prioritize them in curation
- include them in the next dataset and evaluation pack
Why this workflow matters
It creates the compounding loop between operations and learning that differentiates a robust robotics platform from a training tool.
6. Maintenance Escalation
This workflow keeps operations and service work integrated.
- open case from telemetry anomaly, failed update, or operator report
- attach rollout, telemetry, and robot context
- route to specialist queue
- resolve, verify, and close
- promote recurring issues into release or data follow-up when necessary
Why this workflow matters
Maintenance work is one of the main places where real-world failure context becomes durable operational knowledge.
7. Regional Compliance And Access Review
This workflow governs sensitive actions that may have jurisdictional implications.
- review data export or cross-border access request
- validate residency and customer policy
- require approval for remote support or teleoperation exceptions where needed
- record the acting identity, rationale, scope, and outcome
Why this workflow matters
Global platforms need a way to operationalize residency, support-access, and export controls without handling them outside the product.
Workflow Controls That Matter Everywhere
Human gates
Review, release approval, rollout sign-off, teleoperation authority, and maintenance escalation must be expressible in the same workflow system.
Retry and rollback logic
Failures should not collapse into silent job death. They should retry with policy, branch to fallback steps, or escalate to humans with preserved context.
Event-driven triggers
Dataset finalization, candidate promotion, canary regression, update drift, or intervention spikes should all be able to trigger downstream workflows automatically.
Durable execution history
Every workflow run should preserve its inputs, outputs, actor identity, timing, and failure context so teams can reason about what happened later.
Why These Patterns Matter
The individual platform surfaces are powerful, but the platform becomes materially stronger when those surfaces are composed into explicit, inspectable workflows. That is what turns lifecycle adjacency into lifecycle integration.