Anchor to the exact site
Blueprint starts from one real facility and one real workflow, so geometry, constraints, and task context are not guesses.
How It Works
Exact-site world models beat generic simulation when deployment gets specific.
Blueprint starts from one real customer site and adds controlled variation around it for fine-tuning, exact-site data generation, and checkpoint comparison. Research like MIT's RialTo helps explain why that matters: target-environment training tends to beat generic scenes when the deployment question depends on one exact facility.
The operating idea
Exact site plus controlled variation is the training loop that works.
The exact site is a strong anchor, but the biggest lift comes when teams can rerun the task under realistic variations and feed those results back into their training stack.
Branch realistic variations
Once the site is grounded, teams can change lighting, clutter, start states, and other conditions without losing the real-site anchor.
Run, score, and compare
Teams can rerun policies, inspect failures, compare checkpoints, and decide what to export back into training on the same exact environment.
Export data back into the stack
The walkthrough alone is not the useful output. The rollout data, failure cases, and site-grounded evidence are what you feed back into training and deployment decisions.
Comparison
Where Blueprint fits in your training pipeline
Generic simulation is still useful. Final on-site validation is still necessary. Blueprint sits in the middle, where one exact site can answer real deployment questions before the expensive part starts.
Generic simulation
Best for
Broad pretraining and early iteration
Watch-out
Customer-specific geometry, task semantics, and failure modes
Exact site only
Best for
Grounding to the real geometry and checking basic fit
Watch-out
Fine-tuning and edge-case probing if the environment stays static
Exact site plus controlled variation (Blueprint)
Best for
Policy fine-tuning, site-specific training data, and release comparison before deployment
Watch-out
Still requires final on-site safety validation
Common jobs
What teams train and ship with this
Once a team has one specific customer site, these are the jobs they actually run against it.
Pre-deployment training
Use the real site as the training anchor before the first travel-heavy customer week starts.
Policy fine-tuning
Adapt the stack around the place the robot actually needs to work instead of tuning against a generic scene.
Regression checks
Compare releases on the same site so weak updates show up before they reach the field.
Customer readiness
Show operators, buyers, and internal teams the exact site and the expected robot behavior in the same surface.
Example workflows
How this looks across different facility types
These examples show how Blueprint packages and hosted evaluation would be framed for different industries. Named customer references will be added as they become available.
Illustrative preview
Kitchen articulation lane
Illustrative package framing for teams evaluating ovens, racks, drawers, and handoff surfaces in a kitchen lane.
- Show the real lane before talking about automation claims
- Explain what artifacts ground the buyer's own stack
- Show the hosted exports without implying customer proof that is not public
Illustrative preview
Warehouse tote lane
Illustrative hosted-review example for a warehouse lane where the buyer cares about reruns, exports, and failure review on one site.
- Show the setup-to-rerun-to-export loop
- Keep the same facility attached to every artifact
- Make the runtime question legible before travel
Illustrative preview
Retail shelf grasp-place lane
Illustrative retail example for teams that need shelf-level context, package notes, and export framing before the real site visit.
- Label the lane, the shelf logic, and the export types clearly
- Show where privacy and rights would be attached
- Keep the proof honest while real customer references are still private
Start with one real site and one deployment question.
That is enough to decide whether you need the package, hosted evaluation, or a custom engagement. The rest of the workflow gets much cleaner once the site is grounded.