Launch App
Back

Arctivar: The Full Walkthrough

Every move you make becomes training data. A full walkthrough of Arctivar — how to find a task, run it, correct a policy, and turn your work into points.

Every move you make becomes training data. A full walkthrough of Arctivar — how to find a task, run it, correct a policy, and turn your work into points.

Background

Robot learning isn't limited by algorithms — it's limited by a shortage of high-quality, diverse training data at scale.

That's what Arctivar is for. You drive a robot in your browser, and every move you make becomes the data that trains it.

The Three Stages of a Task

Every task moves through three stages:

  1. Pre-training — the model doesn't exist yet. You drive the arm from scratch, and each valid run is one demonstration.
  2. Training — once all twelve hundred runs are aggregated, they're turned into a small policy. The task stays visible but stops accepting runs.
  3. Post-training — the policy drives itself. Its success rate isn't guaranteed, so the moment it's about to fail, you take over, fix what's missing, and hand control back.

You can work on the first and the third — and the difference between them is the key to understanding Arctivar:

  • In Pre-training, you show the robot how.
  • In Post-training, you capture exactly what it can't handle yet, which is the most valuable data on the platform.

What One Full Contribution Looks Like

  1. Find a task that still has openings
  2. Run it — as a demonstration, or as a correction
  3. Sign the run on-chain
  4. Earn points for what you produced

Now, let's see how to put it into practice.

Getting In

  • Log in with Privy. A social account or your own wallet both work, and the system creates a wallet for you.
  • Keep a little ETH on Base. Signing data on-chain costs gas, and a little lasts a long time.
  • Open the getting-started guide. The lightbulb in the bottom-right corner holds a detailed guide in four sections — find a task, pre-training, post-training, sign and verify — with screenshots.

Finding a Task

  • Read Remaining slots first. It's how many openings are left in this stage. It counts entries, not people — when it reaches zero, the task can no longer be run.
  • Filter when the list gets long. By difficulty, scenario, skill, or stage.
  • Spread yourself around. Varied data is what matters most — the platform rewards breadth across scenarios and skills, not just volume.

Running a Pre-training Task

  • Read the Goal and Steps before you move. Steps are an ordered list that lights up as you go.
  • Check the Goal Reference. Bottom-right, it shows the finished state. Layouts are randomized each run, so treat it as a reference, not a copy.
  • Use the keyboard, not the drag ball. Dragging the blue ball works, but it makes messy trajectories. Arrows plus E and D translate, Q W A S Z X rotate, SPACE toggles the gripper.
  • Set up the camera. Zoom, orbit, and use the two wrist cameras until you can clearly see the gap between the gripper and the object. Most people skip this — and that's a mistake.
  • Double-click an object to select it, and the arm closest to it comes over on its own.
  • Slow down before contact. Stop just short, throttle before you translate, or you'll knock things over.
  • Use checkpoints. N saves the current state, B restores it — so a mistake doesn't force you to redo the whole run.
  • Go fast and clean. There's a timer up top: a quicker, more decisive run means a more direct trajectory and less wasted motion, and the score influences how many points you earn. Place the object where the goal says is enough — the system judges completion automatically, with no submit button to press.

Running a Post-training Task

  • Check Policy success first, so you know how often it makes it on its own.
  • Watch the Takeover Budget. You get thirty seconds of manual control for the entire session. Go over it and the run fails outright — and the shorter the takeover, the higher the score.
  • Don't grab control at the first stumble. The policy getting through a hard moment on its own is exactly what the system wants to verify.
  • Time the takeover. Too early wastes budget; too late and the failure already happened. Step in when it's clearly stuck — grabbing at nothing, over and over.
  • Correct only what's missing. Get the arm back into a state the policy can handle and press Return to Policy. You don't have to finish the task for it.
  • Reset instead of rescuing a lost run. If the budget is nearly gone, the score is gone with it; Reset re-randomizes the layout and gives you a fresh start.
  • Undo and Redo work like they do in a document, if a correction goes wrong.

The point of all of it: less takeover time, a clear view, a quick handback.

After the Run

  • Sign every run on-chain — it doesn't count otherwise. It's a transaction on Base recording the task ID, the data ID, and the user's address, which you can look up in Explore.
  • Track your points in Portfolio → Points. They're grouped into periods called Epochs, split into Task points from your own runs and Referral points from what your invitees produce. Points are not a token: not transferable, not redeemable for cash.
  • Check Alliance for partner tasks. Both pre-training and post-training, run with partner projects such as BitRobot — same way of working, some need an access code, and their points stack on top of Arctivar points.
  • See where you stand on the Leaderboard — trajectories, referrals, and points, each with time windows.

That's the whole chain: find a task with room, run it, sign it on-chain. It trains a policy, you come back and correct that policy — and every turn of that loop moves the robot forward.