Tancho: the drone decides where to look next

An on-board model that reads a disaster scene, names the view that is still missing, and flies the drone there. The model is a fine-tune of NVIDIA Cosmos 3 Edge; training and the frozen evaluation run on Nebius AI Cloud, and the ground-side tasking step runs on Nebius Token Factory with NVIDIA Nemotron.

Open weights: v13s/tancho, revision 1d2b42a3e00542a63bbaf216295c7d119c706b2b (2.44B parameters, BF16).

Demo video (3 minutes): https://youtu.be/3Kl3QwWzflA

The loop

1. The aircraft captures a frame.
2. Reasoner (Cosmos 3 Edge fine-tune) reads it against the observation goal and reports whether the goal is met, or what is missing.
3. The viewpoint readout proposes the next viewpoint as a direction and a distance, inside a volume approved before takeoff.
4. A deterministic compiler turns that viewpoint into a flight path and checks it against the approved volume, the altitude ceiling, obstacle clearance, dwell time and the pixels required on the target. A proposal that breaks a limit is rejected before the aircraft moves.
5. The aircraft flies, captures again, and the Reasoner judges the new frame.

Nothing outside step 2 and 3 is produced by a language model. Detection, geometry and limit checking are deterministic, so they can be audited.

Where NVIDIA and Nebius are used

Step	What runs	Where
Parent model	NVIDIA Cosmos 3 Edge, through NVIDIA cosmos-framework	-
Training, frozen evaluation	the fine-tune and its 104-case evaluation	Nebius AI Cloud Serverless Jobs, gpu-h100-sxm 1gpu-16vcpu-200gb, image pinned by digest in Nebius Container Registry, parent weights on a Nebius compute filesystem
Tasking to observation goal	NVIDIA Nemotron 3.5 Lightning	Nebius Token Factory
On-board inference	the published fine-tune	Jetson-class board (measurement pending)

Install

git clone <this repository> && cd tancho-nebius
uv venv && uv pip install torch transformers pillow

The evaluation also needs NVIDIA cosmos-framework at revision 96303bb0bdd14d9efa18f24d8ef98c7a0bfb8412, which supplies the Cosmos 3 Edge model class and processor:

git clone https://github.com/NVIDIA/cosmos-framework.git
git -C cosmos-framework checkout 96303bb0bdd14d9efa18f24d8ef98c7a0bfb8412
uv sync --frozen --all-extras --group=cu130-train --directory cosmos-framework

Run the ground-side tasking step (Nebius Token Factory)

Put a Token Factory API key in ~/.config/tancho/token_factory_api_key (mode 600) or in NEBIUS_TOKEN_FACTORY_API_KEY. Never pass a key as a command-line argument.

python3 tools/demo_tasking.py

That is the step shown in the demo video: two taskings, each turned into one observation goal and checked against the catalog. One call on its own is three lines:

python3 -c "
from tancho.tasking_to_goal import request, parse_goal, validate_goal
r = request('Check whether the road under the slope is still passable.',
            'nvidia/Nemotron-3_5-Lightning')
goal = parse_goal(r['raw']); print(goal, validate_goal(goal))"

The model may only choose from the closed vocabulary of the mission profile catalog in tancho/mission_profiles.py; validate_goal returns every violation, and a goal with any violation never reaches the aircraft.

Reproduce the model comparison

python3 training/evaluate_tasking_models.py \
  --model nvidia/Nemotron-3_5-Lightning \
  --model nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B \
  --model nvidia/nemotron-3-super-120b-a12b

The 16 cases in training/tasking-goal-cases-v1.json were frozen before any model was called. Results are in docs/nemotron-tasking-comparison.md.

Run the frozen evaluation on Nebius AI Cloud

deploy/nebius_exhibition/prepare_evaluation_job.py creates a Serverless Job that downloads the published revision, clones the pinned cosmos-framework, clones this repository at the commit it pins and checks the frozen pack's manifest hash, then evaluates all 104 cases on one H100. It validates with --dry-run by default and only creates a job with --execute.

Three things come from your own Nebius account, so the script reads them from the environment rather than carrying ours:

export NEBIUS_PROJECT_ID=project-...          # the project the job is created in
export NEBIUS_FILESYSTEM_ID=computefilesystem-...  # holds the Cosmos 3 Edge parent weights
export NEBIUS_IMAGE_REF=cr.<region>.nebius.cloud/<registry>/<image>@sha256:...

python3 deploy/nebius_exhibition/prepare_evaluation_job.py            # dry run
python3 deploy/nebius_exhibition/prepare_evaluation_job.py --execute  # creates the job

The job prints one TANCHO_EVAL line with the score. Our run scored 104 of 104, with exact_accuracy 1.0 and schema_valid_rate 1.0, in 313 seconds.

Before you publish anything from this tree

python3 tools/scan_public.py .

It fails on client names, cloud resource ids, local absolute paths and credentials.

License

MIT for this code. The published weights are a derivative of NVIDIA Cosmos 3 Edge and carry the parent model's license; see the model card at v13s/tancho.