Route Recovery

• 

  Actual synthetic grid simulation: ten moves and one replan. No robot, physical safety validation, or cloud execution.

• 

  Mobile synthetic route recovery with audited movement. Live NVIDIA model and Nebius cloud-job execution remain unverified.

The problem

When a delivery route becomes blocked, a recovery suggestion needs a visible state, constrained actions and an inspectable outcome. Route Recovery lets an operator test that process in a fully known synthetic grid.

The product

Change obstacles or restrictions, inspect a recovery suggestion and explicitly apply it. NVIDIA Nemotron Lightning (nvidia/Nemotron-3_5-Lightning) on Nebius Token Factory selects follow, hold or request operator clearance for the exact reviewed state. A deterministic breadth-first planner validates movement; the model cannot invent coordinates, clear an obstacle or declare success. A changed map invalidates the old decision. The export preserves the actual action history.

Sponsor technology and build

The Node.js and browser implementation separates model advice from constrained execution. A real Tavily search retrieves navigation references, and selected passages feed a separate Nemotron evaluation checklist. Web content never grants movement clearance.

A reviewable Nebius AI Cloud Serverless Jobs bundle is implemented to evaluate the same state and policy. No Cloud job has run: verified funded access is still unavailable. Preparing a bundle is not cloud execution. Deepgram Aura-2 Thalia generated narration audio that passed technical checks; listening review and the final module video are pending.

What is verified

One live Nemotron decision was applied to the synthetic detour scenario. Ten audited grid moves reached the goal with zero constraint violations and a valid saved history. This establishes that scenario only. The standalone full suite passed 82 tests; later focused checks passed 17 research units and two API/browser checks. Counts overlap, and local fixtures remain distinct from live inference.

The latest research output quotes relevant pause/replanning options but assumes Nav2 and XML details that this product does not implement. Original outputs are preserved and a separate assistant review narrows the comparisons. The product uses no robot hardware, sensors, physics engine, Isaac Sim or Nav2 runtime. It does not establish physical safety. Physical AI qualification remains unverified while the real Jobs proof and required module footage are unfinished.

Run and current access

Repository. As checked on 22 October 2026, this repository is private and its remote revision predates the validated local release. Publication and judge access remain pending; the repository link is not yet a public reproduction path.

The commands below apply to the prepared standalone release after access to that revision is available. Node.js 22+ is required. Live sponsor actions need private environment configuration and explicitly reviewed provider access; local examples work without provider keys.

npm ci
npm start

Open http://127.0.0.1:4190/route/. For local fixture tests, run npx playwright install chromium and npm test without provider credentials. Public judge access, the completed Cloud evaluation and the final video remain pending.

AI tools and development

Codex assisted with implementation, interface refinement, test authoring, debugging and independent review. NVIDIA Nemotron performed the bounded runtime tasks described above; Tavily supplied search results, and Deepgram generated narration. AI output was checked against saved state, validation rules and actual execution evidence. This description does not claim that another assistant's separate hackathon projects are part of this entry.

Built With

• css
• deepgram
• html
• javascript
• nebius-token-factory
• node.js
• nvidia-nemotron-lightning
• playwright
• tavily