ubr-physical-ai/rescue-brain-pipeline

Rescue-Brain Pipeline

UB Robotics (team UBR Stack) working plan for the NVIDIA Open Models Codefest 2026: a reusable data → train → edge → sim → field pipeline with a measured gate at every stage, demonstrated on search and rescue and re-run as a transfer test on early fire/smoke ignition.

Plan v3 (12 Sep): off-the-shelf detector baseline (B0/B1) is the reference our own model must beat; gates are a reporting order, not a build order; and the edge brain is decided.

Day 2–3 (10–11 Sep): the 4-bit edge reasoner was built, exported and executed on an 8 GB Orin Nano — 2.24 GiB of engines, ~4.83 GB full-VLM peak, 52.8 tok/s decode, 169 INT4 nodes verified. On 24 real frames it finds 9–10 of 12 targets where the incumbent finds 0–1 (paired p = 0.002–0.004), and never once emits the output schema, so the contract moved into a parser and the geometry into code. Decision (11 Sep): Cosmos3-Edge takes both jobs — perception and the operator interface (English commands 13/15, 5/5 on stop synonyms) — with Gemma kept as a measured text-only fallback.

Pages

Start at `index.html` — the plan, with an index into the evidence pages below.

• —`index.html` — the pipeline: five gates plus the transfer test, where each Codefest criterion is answered, and the current state of each gate.
• —`progress.html` — Three Days In: what 9–11 September produced. Measured results with their N, the results that changed the plan, and an honest list of what is not settled.
• —`blockers.html` — Blocker Board: what is open, what it blocks, and what unblocks it.
• —`pointers.html` — Pointers Followed: the five resources suggested at kickoff and what came of each — adopted, explored, deferred, and one gap.
• —`bakeoff.html` — Cosmos vs Gemma: the two edge VLMs on real frames — who perceives, who writes the contract, who decides.
• —`testing.html` — the Test Ladder: how a model is judged, L0 through L3.
• —`baseline.html` — YOLO vs PeopleNet: the off-the-shelf detector reference (B0/B1) our own model must beat.
• —`datasets.html` — Datasets: what we built and published, the sources we used, and the data challenges still open.
• —`contributions.html` — Contributions: public models, datasets and code, and the issues and patches sent upstream to NVIDIA's edge toolchain.

All pages are static renders of the team's evaluation reports; the reports are authoritative where they disagree.