HUGIN LEARNINGby Project Hugin
GPU Owner GuideLogin

Robotics · Drones · Automation

Hugin Learning

The learning platform for robotics, drones and automation.

Teach your robots, drones and machines how to act and what to look for — with almost no data. Everything else you need is a Hugin Learning feature.

>start hugin learningExplore ↓
  • Encrypted by default
  • Price before you start
  • Your model, on your machines

Hugin Learning · Control behaviour

Learns how to act

Robots, drones and machines discover the right behaviour by trial and error.

No demonstrations or labelled data needed. Pick a built-in task or connect your own simulator — Hugin Learning finds the policy, and you download it to run on your robot.

  • MuJoCo · Box2DGymnasium tasks built in
  • Your simulatorIsaac Sim · Gazebo · any Gym env
  • No labelled dataLearns by trial and error
  • Policy fileReproducible · yours to deploy
Robot control policy · from €2.52Try it →

Hugin Learning · Find objects

Learns what to look for

Show it a few photos — it finds the object in new photos and videos.

Three photos of an apple — or of your own part, tool or product — are enough to start. It shows exactly where the object is, tells you when it isn’t there, and runs on your own cameras, robots and computers.

  • 3 photosEnough to start
  • Photos & videoFinds it in new footage
  • Where — or not thereMarks it · says when it is missing
  • Runs anywhereYour cameras · robots · PCs
Object finder · from €1.28Try it →

Features

Everything is a Hugin Learning feature

One learning platform at the centre. Perception training, simulation, protected GPU compute and verification all work for it — so your machines learn faster, cheaper and in private.

The core

Hugin Learning

Learns how to act and what to look for — with almost no data.

Start learning →

Feature 01

Perception training

YOLOv8 detection, segmentation, pose and oriented boxes, or ResNet / EfficientNet / ConvNeXt classifiers on your own images.

Feature 02

Simulation & validation

Benchmark model versions on the same test set, run scenario sweeps and repeat seeded experiments.

Feature 03

Batch inference

Run your trained model over large image sets — inspection, mapping and QA at scale.

Feature 04

Distributed GPU training

Every job is matched to the right GPUs automatically and recovers if a machine drops out.

Feature 05

Hugin Shield+

Included on every job: your data is encrypted and split, so no machine ever sees the whole.

Feature 06

Hugin Vault

For your most sensitive data — runs only on sealed hardware that even its owner cannot read.

Feature 07

Verified results

Every result is independently re-checked before it reaches you — a model you can trust.

Feature 08

Language models for automation

Fine-tune open LLMs (LLaMA, Mistral, Qwen) on your procedures for operator assistants.

How it works

From task to trained skill

Six steps — with the price shown first.

  1. 01Pick what to learnA task in your simulator, a few photos of an object, or your labelled dataset.
  2. 02See the priceIn euros, with an upper bound, before the job runs.
  3. 03Matched automaticallyYour job goes to the right GPUs across the network.
  4. 04Learns securelyUnder Hugin Shield+ or Vault, recovering if anything is interrupted.
  5. 05Quality verificationEvery result is checked before it reaches you.
  6. 06Download & deployRun the policy file or model weights on your own robots, drones and machines.

Hugin Learning · Protection

Hugin Shield+

Protection built into every Hugin Learning job — at no extra cost.

Most GPU networks ship your raw data to strangers’ machines. Hugin doesn’t: your data is encrypted and split, so no machine ever sees the whole.

  • Private by defaultProtected the moment you upload
  • Yours aloneNo one sees your full data
  • Trusted machinesVerified, reliable partners
  • Peace of mindNo setup · no risk
Included on every jobStart a job →

Hugin Learning · Confidential computing

Hugin Vault

Maximum confidentiality — your data stays invisible, even to the computer that trains it.

For your most sensitive data: jobs run only on hardware-isolated machines that are verified before they touch your data. The machine’s owner can never see it.

  • Sealed spaceOpened only inside a protected area
  • Invisible to ownersNever readable by the host
  • Proven genuineEvery machine verified first
  • Sensitive dataHealth · finance · personal records
For your most sensitive workChoose Vault →

Pricing

Hugin Learning pricing

Pay per job. See the price — with an upper bound — before anything runs. 1 HU = €0.21.

JobShield+Vault
Robot control policyHugin Learning · 60 generations€2.52€6.55
Object finder from a few photosHugin Learning · 3 photos · compact€1.28€3.32
Defect-detection modelYOLOv8s · object detection · 50 epochs€31.50€81.90
Drone / aerial detector, lightweightYOLOv8n · object detection · 50 epochs€15.75€40.95
Image classifierResNet-18 · 30 epochs€14.18€36.85
Simulation sweep10 scenarios€63.00€163.80
Batch inferenceYour trained model over an image set€2.10€5.46

Figures come from the same cost calculator you see in the dashboard. Shield+ is included at the standard rate; Vault runs on hardware-isolated machines at 2.6× the standard rate. You are charged the estimate shown when you queue the job.

Start with Hugin Learning →Talk to us first

FAQ

Questions

What is Hugin Learning?
Hugin Learning is our learning platform for robotics, drones and automation. It teaches robots, drones and machines how to act — by trial and error, in simulation — and what to look for, from just a few photos. Perception training, simulation, protection and our distributed GPU network are all features of it, so a whole project runs in one place.
What is Project Huginn?
Project Huginn is the home of Hugin Learning — the learning platform for Physical AI: the robots, drones and industrial automation that work in the real world. Teams teach control behaviour, train perception models, validate and deploy, without owning a data center. Built by SMware ApS, a Startup Denmark approved company.
Which frameworks and simulators are supported?
Vision: PyTorch and Ultralytics YOLOv8 (object detection, segmentation, pose / keypoints, oriented bounding boxes, classification) and ResNet, EfficientNet, MobileNet, ConvNeXt and VGG classifiers. Control policies: Gymnasium, MuJoCo (Reacher, Hopper, HalfCheetah, Walker2d, Ant, Humanoid) and Box2D environments are built in; Isaac Sim, Gazebo or your own simulator connect through a Gymnasium-compatible interface. Sensor and telemetry data: CSV and JSONL.
What can I build for robotics, drones, and automation?
With Hugin Learning: control policies for balancing, reaching, locomotion and navigation, and object finders that learn from a few photos. With its other features: defect and obstacle detectors, aerial-imagery models, keypoint models for manipulation and image classifiers for inspection; validation and benchmarking of model versions, simulation sweeps, repeatable seeded experiments, and batch inference with your trained model. You can also fine-tune open language models for operator assistants.
How does Hugin Learning train robot and drone policies?
Pick a built-in Gymnasium, MuJoCo or Box2D task, or connect your own simulator through a Gymnasium interface. Hugin Learning improves the policy by repeated trial runs spread across the GPU network — no labelled data needed — and every run is reproducible. When it finishes, you download the policy file.
Can Hugin Learning work with just a few photos?
Yes. Name the object and show it a few photos — three is enough to start. It then finds that object in new photos and videos, marks where it is, and tells you when it is not there. You download the model and run it on your own robots, cameras or computers. More photos make it more reliable.
What do I upload?
For YOLOv8 detection, segmentation, pose or oriented boxes: a zip with your images, labels and data.yaml in the standard Ultralytics layout. For classification: images in one folder per class. For sensor data: a CSV or JSONL file with a label column. For custom policies: your Gymnasium environment.
How does billing work?
You pay per job in Hugin Units (1 HU = €0.21). Before any job starts you see its estimated cost with an upper bound, and that is what you are charged when you queue it. See typical job costs under Pricing.
Are my datasets and model IP secure?
Yes, and you choose the level per job. Hugin Shield+ is included on every job: your data is encrypted and split so no single machine ever sees the whole dataset. For the most sensitive work, Hugin Vault runs the job only on hardware-isolated machines, where even the machine owner cannot read your data. Your datasets, policies and trained models stay private to you.
Can I download my trained model and deploy it to a robot or drone?
Yes. When a job finishes you download the weights — YOLOv8 and PyTorch models as .pt files, Hugin Learning policies as a policy file — and run them on your robot, drone or edge device with your own runtime. Your models are yours.
Can I talk to someone before signing up?
Yes. Email the team or book a call from the contact section — we are happy to look at your use case and data format first.

For GPU providers

Can I earn by sharing my GPU?
Yes. Your spare GPUs can power Hugin Learning: connect them to the network and earn as they process robotics, vision and automation jobs. Earnings are paid for verified, completed work.
How do I connect my GPU?
Sign in, choose the GPU Owner role, open the GPU Agent page in your dashboard and follow the guided setup. The agent detects your GPU automatically and starts taking jobs. The full guide is on our How to Share page.
Which GPUs can join the network?
Most modern NVIDIA GPUs, from workstation cards to data-center GPUs. Each job is matched to machines with enough memory for it; machines with confidential-computing hardware can also take Hugin Vault jobs.

Talk to us

Tell us what your robots, drones or machines need to learn — or email info@projecthuginn.com