UBR Stack · NVIDIA Open Models Codefest 2026 · working plan v3 · 12 Sep
Rescue-Brain Pipeline
The deliverable is an autonomous search-and-rescue robot: it searches, decides what to approach, and reports on its own — and keeps working when the link is cut. We build and prove it through a reusable data → train → edge → sim → field pipeline with a measured number at every gate, demonstrated on rescue and re-run on a second target to show the method transfers. Every week ends with a number that did not exist the week before.
How the organisers pick the GTC Berlin five — and where each criterion is answered
Progress
One gate metric per week, stage by stage
Relevance
Civil protection & SAR — a global public-sector problem, made sharply concrete by the 2026 disaster year
Uniqueness
Real rover, real judge set, an off-the-shelf baseline our own models must beat, a transfer test, and fixes sent upstream to NVIDIA's edge toolchain
Realistic use-case
Synthetic ≠ field accuracy; the physical rover is the final gate
Technology
Isaac Sim · Cosmos Transfer · Cosmos3-Edge · TAO · TensorRT Edge-LLM · RAPIDS · NeMo Curator
Why now — the 2026 field
2026 has been an exceptionally destructive year for exactly these hazards. Earthquakes killed about
7,100 people worldwide — Venezuela (M7.5, June, 6,500+), Colombia's Chocó (M7.4, August, 331),
Indonesia's Flores (M7.7) and the Philippines' Mindanao (M7.8). In late August a glacier collapse
above Langtang Lirung sent a flood across the Nepal–Tibet border: 1,385+ dead, 5,400+ missing,
~1.6 million affected. And Europe logged its fourth-worst wildfire season since 2012 —
489,826 ha burned by early August, Spain and France hardest hit, Portugal 47,776 ha — the
same hazard behind our transfer test. Each is a ground search-and-rescue problem in comms- and
GPS-degraded terrain, which is what this pipeline is built for.
MECHANISMHow data becomes a measured brain
Three sources plus subsets of NVIDIA PhysicalAI datasets pass a single QC gate before the B300 touches them; the judge set bypasses training entirely and feeds only the scorer. Sim and real harnesses run the same placements and write the same log to one scorer. For perception that makes the sim-to-real gap a subtraction — same model, different imagery. For missions it does not: the sim brain runs on a workstation GPU and the rover brain on an Orin, so we compare the two directly rather than treating one as the other's baseline. What the scorer finds missing goes back to lw-lab1 as the next render run.
Score the hybrid set before training anything: coverage (distance bins incl. past the 20 m wall — a body of a few dozen pixels at the detector's input resolution — pose classes with hard negatives, lighting, smoke), label consistency (do the masks and boxes still agree after Transfer?), and the sim-to-real gap: the same detector scored on sim and on real frames. Build the judge set: a mannequin capture on a taped floor supplies the lying positives with measured distances; the 474-frame past-mission slice supplies the negatives.
Gate metric
Coverage table · label-consistency % · off-the-shelf vs ours on real frames · recall per distance bucket · QC report structured by ISO/IEC 5259 dimensions
NVIDIA tech
Isaac ReplicatorCosmos Transfer 2.5RAPIDS cuDF Replicator on lw-lab1 (RTX); Transfer on the B300. Measured: segmentation-only control is under-constrained — a single-instance mask covers ~1 % of the frame, labels survive ~13 % and 0 % keep the hi-vis vest. A static-subject control reaches ~0.90, so the mechanism works when adequately supplied: moving to multi-control (edge + seg + depth), not tuning strength.
2
Weeks 1–2
Train
Detector sweep: YOLO26n vs PeopleNet × 640 / 1280 × ±Transfer variants × ±hard-negative pseudo-labels (teacher = Cosmos3-Super or an open-vocabulary grounder — and treated as a noisy label, not ground truth: a prompt change alone swung its box count 3× on identical frames). Any bounded SFT of Cosmos3-Edge currently targets the operator-language side (tool-calls, an UNKNOWN reply); the perception false-alarm rate is under re-measurement on the 474-frame slice. Ablation against an off-the-shelf reference: B0 pretrained COCO person and B1 PeopleNet as shipped, each plus a fixed posture rule, then D1 sim-trained → D2 +Transfer → D3 +real. The headline claim is the step from B to D, not the steps among the Ds.
Gate metric
Confidence vs distance, median + IQR over N = 20 placements across 8 buckets, both models · false-alarm rate per threshold on the 474 real mission frames · COCO mAP · the posture rule's false-alarm rate on non-lying poses, with lying recall held. External: AI City test splits (SmartSpaces, Spatial-QA) + SARD lying-person recall
NVIDIA tech
TAONeMo / LoRAB300 Sweeps as background processes in one srun
3
Week 2
Edge
Distil the winning detector into the nano the rover actually runs. Build the TensorRT Edge-LLM INT4 engine for Cosmos3-Edge on the Orin Nano. Orin is FP16 / INT8 / INT4 only.
Gate metric
Met, 11 Sep. INT4-AWQ on the Orin: 2.24 GiB engines, ~4.83 GB full-VLM peak, 52.8 tok/s decode, 6.0 / 10.1 s per look — under ≤ 5 GB and 5× the ≥ 10 tok/s floor. Still open: watts, p95, co-resident memory with the voice stack, and the SM103↔Orin engine-equivalence gate.
NVIDIA tech
TensorRTTensorRT Edge-LLMJetPack 7.2
4
Week 3
Sim rover
Closed-loop mission in Isaac Sim on lw-lab1: briefing → search → detect → verify → mark → re-plan → report, with scripted comms loss. N seeded scenarios. This is where the agent is measured, not only the models — and three of its decisions get their own numbers: decide to approach (a casualty in view that the rover chose not to close on), decide to report and the harder decide not to, and behaviour under comms loss. The brain is decided (11 Sep): Cosmos3-Edge does both perception and the operator interface; geometry stays deterministic in code, with a stop matcher ahead of the model. The mission harness runs, with one scored shakedown so far.
Gate metric
SubT rule: found within 1 m (ours) and 5 m of ground truth — the report's position, not the rover's · time to report · false reports · VLM calls · distance driven · Success / SPL / distance-to-goal (ObjectNav). N = 20 paired placements per arm; McNemar on the binary metrics, Wilcoxon on the continuous ones, one primary metric declared before the runs; median + IQR, bootstrap CI. Baseline today: one mission, 822 s, 56 VLM looks
NVIDIA tech
Isaac Sim 6.0Nav2 Depends on the planner track's closed loop
5
Weeks 3–4
Physical rover
Same mission on the bench with a mannequin. The network cable is pulled mid-run: the rover keeps searching; the field report arrives on reconnect. Real rover-camera frames are the final generalisation gate for every model.
Gate metric
Same metrics as stage 4 → the mission sim-to-real gap · report on reconnect · video · ISO 18646-2 navigation tests on the same floor (pose repeatability, obstacle detection/avoidance, path deviation) · E2853-style scoring sheet
NVIDIA tech
Jetson Orin Nano 8 GB
T
Week 4 · 3 days
Transfer test
Re-run stages 1–4 on a second target with the same rover, sensors, scenes and tools: early fire / smoke ignition — UBR's founding problem. Only the target class and the assess_detection prompt change. No physical-rover stage. Capped at one column and three days, with no scope growth — it is the reusability proof, not a second project. What the cap frees goes into stage 4/5 autonomy.
Gate metric
Second column of the same table: same pipeline, new target, X days, judge-set P/R Y
NVIDIA tech
Identical stack — that is the point
COMPARABILITYWhat each number is compared against outside the lab
Layer
External reference
Status
We report
Navigation
ISO 18646-2:2024 — service-robot navigation test methods
run as specified
pose repeatability, obstacle detection/avoidance, path deviation — rover and sim, same geometry
No single standard covers "ground rover finds a prone casualty indoors". We run what can be run, report in the format of the rest, and never say "certified".
RULESWhat keeps the numbers honest
Never train on the 60 frozen frames or on the judge set. New training and validation sets for every experiment.
Real rover-camera frames are the final generalisation gate for every model. Synthetic results are diagnostics, not field accuracy.
Ablation by source is mandatory — sim-only / sim+Transfer / sim+Transfer+real. It is the proof that the hybrid dataset earned its place — the kind of result judges remember.
The transfer test must not touch the rover code path. New tools or new sensors means it is not a transfer test, and it is out of scope.
The baseline is measured, not remembered — and it is not ours. B0 (pretrained COCO person) and B1 (PeopleNet as shipped) are the reference; our sim-trained model is D1, a result rather than a baseline. If B1 + a ten-line posture rule matches D1 on real frames, the synthetic pipeline earned nothing — and we would rather find that out ourselves.
The gates are a reporting order, not a build order. The whole path — frame → detector → VLM → decision → report → scorer — is stood up end to end on ten frames, one sim mission and one rover run before any gate is filled. A gate's number may be a placeholder; the code path it runs through may not.
No claim below the noise floor. Two runs of the same model in different containers agree on ~95% of items (88% on distance), so a smaller gap is not evidence. And a worked example shares no content word with any test item — three contaminations were caught exactly this way.
Keep the agent in front. Detector numbers feed assess_detection; the demo is still the search → verify → comms-cut → report loop.
The decisions are the product. Approach, report, and above all not report, each measured per mission. Any SFT budget now goes to the operator-language side — tool-calls, an UNKNOWN reply, and multilingual capability (demonstrated in European Portuguese) — since 4-bit perception costs nothing once calibrated and the model is no longer asked to emit the contract.
DEPENDENCIESTo confirm this week
Planner track: closed loop in Isaac — running
The L2 mission harness runs on lw-lab1 and has scored one shakedown mission; the paired comparison has not been run. One open call: run arm C's paired placements on jetson1, or accept the workstation engine behind an SM89 equivalence gate.
Answered — decision on the engine
Mannequin capture for the judge set — scheduled for Day 4
It supplies the lying positives with measured distances — lying recall per distance bucket. With the capture on Day 4, those numbers arrive late: until then the 474-frame slice carries the false-alarm side, and the earlier ablation is detection and false-alarm rate, not lying recall by distance.
Day 4 · 30 Sep
lw-lab1 is double-booked — measured render-bound
One RTX 4060 Ti 16 GB serves both the render backlog and the missions, and Isaac needs RT cores so neither can move to the B300. Measured: missions run at rtf 0.34 and staging costs ~4× a mission, so placement ordering, not the model, drives the schedule.
Sequence, do not assume
Licence of Isaac-rendered output
Whether rendering with NVIDIA SimReady assets constrains the licence of the published images. The one public forum thread asking this was closed with "contact Enterprise Sales", so it needs an answer from inside NVIDIA.
Asked 11 Sep
CALENDARSessions and exit criteria (Lisbon time)
When
Session
Stage
Exit criterion
Wed 09 Sep
Day 1 kickoff — done
—
Live GCS + rover demo to mentors; cluster access, 1× B300
09 – 16 Sep
async
1 → 2
Judge set captured; B0/B1 off-the-shelf reference measured on real frames; coverage table; Transfer strength swept and its limits published; edge INT4 artefact built; 10 edge tools stubbed
Wed 16 Sep 09:00–16:00
Day 2 scrum
2
Ablation table B0/B1 → D1 on real frames; the edge brain decided with its INT4 numbers in hand; briefing → edge agent → field report on a recorded scene
16 – 23 Sep
async
3 → 4
Edge INT4 gate already met (see stage 3); this week: SM103↔Orin engine-equivalence gate, k per range bucket on the Orin, co-resident memory, and the 20-placement L2 comparison
Wed 23 Sep 09:00–16:00
Day 3 scrum
4 + 5
Closed-loop mission in Isaac (N scenarios, median/IQR); bench mission with cable pull
23 – 30 Sep
async
T
Same pipeline on fire/smoke ignition, stages 1–4, three days; second column of the table
Wed 30 Sep 09:00–16:00
Day 4 scrum
freeze
Dry run < 5 min; backup video; both columns of the gate table filled
Wed 07 Oct 10:00–16:00
Day 5 final
—
Present
Tue 20 Oct
GTC Berlin
—
If top 5
UB Robotics · team UBR Stack · NVIDIA Open Models Codefest 2026