A modern browser (Chrome, Firefox, Safari, or Edge) on a desktop or laptop. No GPU, no downloads, no special hardware. The physics engine — MuJoCo, compiled to WebAssembly — runs locally in your browser.
01Getting Started16
Sign up through Privy — you can use your email, Google account, X (Twitter), or connect an existing wallet. If you don't have a wallet, Privy creates one for you automatically. See Getting Started.
A custodial wallet created for you automatically by Privy when you sign up, so you don't need to bring one of your own. Custodial means the keys are managed for you: there is no seed phrase to write down and nothing to install, and it is not intended as a self-custody wallet for holding assets.
It is what you use to sign each completed task on-chain, which is the step that makes a run count. Your wallet address and balance are visible under your avatar in the top-right corner. If you signed up by connecting an existing wallet, that wallet is used instead.
Signing a run is an on-chain transaction on Base, so it needs a small amount of ETH on Base for gas. Completing and submitting a task itself costs nothing — only signing touches the chain.
In practice the cost is negligible: Base fees run a tiny fraction of a cent, and Arctivar runs gasless campaigns regularly that cover the fee for you outright. See Sign & Verify.
New tasks go up daily at 8:00 PM SGT (12:00 UTC).
Up to 5 times. Each completion is an independent data entry with its own score, but for Points they are not additive — what the task is worth to you is the average across your completions. See How are Points affected when I repeat a task? under Points, Scores & Rewards.
The cap exists because task slots are finite: an account filling them with near-duplicate repeats crowds out the variety the dataset is collected for. See Don't Farm a Single Task.
No. Arctivar Hub requires a desktop or laptop web browser. The simulation runs locally in the browser and the task controls assume a keyboard and mouse, so phones and tablets are not supported.
Failed attempts are not uploaded and don't count against you. You can retry immediately, unless the slots are gone. A run fails in the moment for one of three reasons: the time limit is exceeded (on timed tasks), all slots were claimed while you were playing, or — in post-training tasks only — the 30-second takeover budget runs out before completion.
A run that completes successfully can still be rejected later during verification. That is a different thing — see Why was my trajectory marked as Failed? under Getting Started.
Slots count data entries, not unique users. If a task stage has 1,200 slots, it ends when 1,200 runs have been submitted, regardless of how many people contributed them.
They fill fast because the same user can take several of them (up to the 5-completion cap), because popular tasks draw concurrent submissions, and because a trajectory awaiting verification holds its slot in the meantime.
It is being verified, and verification is not instant. The system has to load the task and the MuJoCo model, replay the entire trajectory, and check both physics consistency and task completion. Long trajectories take proportionally longer. Wait for it to finish — the trajectory then resolves to either signable or failed.
Verification rejected it after upload. The most common causes, in rough order of frequency:
- Physics replay failed
- The task checker failed
- The trajectory was too short
- Invalid trajectory format
- Model or scene mismatch
- A temporary infrastructure issue
The first two account for most new failures. This is separate from failing a run in the moment — see What happens if I fail a task? under Getting Started.
Work through these in order, and stop as soon as it clears:
- Refresh the page.
- Log out, then log back in.
- Create a new Chrome profile and log in to Arctivar from it.
- If you can't create a new profile — or you already use one and the problem comes back — clear your browser data and log in again.
If none of that works, report it in Telegram with the task ID and what you saw.
Each trajectory includes: robot joint states, object poses, control commands, task metadata, and a success/failure flag. In post-training tasks, the recording also marks which segments were executed by the policy and which by you. No personal data is captured — your player ID is anonymized.
Every submitted trajectory is replayed in simulation and checked for physical consistency before acceptance — invalid submissions never reach training. Accepted data then goes through automated cleaning, quality evaluation, and IsaacSim-based augmentation before entering the training pool.
The browser frontend runs MuJoCo compiled to WebAssembly. Physics, collision, and control all run locally on your machine. Rendering is done via Three.js / WebGL. The backend augmentation pipeline runs on IsaacSim with GPU servers.
Runs are recorded on Base. You sign each run from your Portfolio's History tab — a unique Data ID anchors the trajectory, bound to its task ID and your contributor ID for a verifiable provenance trail. The transaction hash is visible in your History once signed. See Sign & Verify.
02Pre-training & Task Controls10
Three, freely mixed: hold and drag the blue ring on the arm to move the gripper along the current view plane, double-click an object to send the arm into a pre-grasp pose automatically, or use the keyboard — arrow keys and E/D translate, Q/W · A/S · Z/X rotate, SPACE works the gripper, and C switches arms on bimanual tasks. Full key table: Controls.
Left-click any item and its name and description appear in the top-left corner of the screen. Use it to confirm you're about to grab the right object before you commit to the grasp.
Raise the sensitivity slider below the viewport — it sets how fast the arm responds to your input.
Follow the instructions in order and complete every step. Make sure each numbered step is checkmarked before moving on to the next one — if you skip required movements and jump to the end state, the run may not be recognized as completed.
N saves the simulator state; B restores the last one you saved. Checkpoints are local to the run and cost you nothing — save one after every completed stage and before anything risky.
Use checkpoints, and use them often. After each successful movement, save a checkpoint with N. If you make a mistake, restore the most recent checkpoint with B and continue from there instead of restarting the whole task. This is what makes long, multi-stage tasks practical to record.
Complete the task as quickly as you can while keeping the motion smooth and deliberate. Pre-training submissions are scored on three things: success, efficiency, and motion smoothness. Frantic corrections and long detours cost you on all three.
In a Pre-training task, you teleoperate the robot to complete the task from scratch — your demonstration teaches the task's base policy. In a Post-training task, that trained policy attempts the task on its own while you supervise; you take over to correct it when it slips, and your correction trains the next iteration. Every task progresses from Pre-training through Training to Post-training — filter by stage in the Hub.
You don't have to declare it — success is detected automatically. The moment the success conditions are met, a Task Completed popup appears with your completion time. That popup is the confirmation.
Your trajectory is packaged and uploaded automatically in the background. No score is shown at that point, though, and none exists yet: the quality score is assigned only after the backend has verified your run by replay, and it appears against the run in your Portfolio's History tab. See Sign & Verify.
While you work, the numbered steps in the left panel checkmark one by one as you clear them, so you can see how far along you are before the run finishes.
Save a checkpoint immediately before the risky moment. Before placing or dropping an item, save one — if the task doesn't verify after you place it, restore the checkpoint and try again from a different angle or position instead of redoing the run from the start.
It also helps to let the arm travel for you: double-click the target object and the arm moves into a pre-grasp pose on its own, so your effort goes into the grasp rather than into crossing free space.
03Post-training4
Error means the policy failed to load, so there is nothing for you to supervise yet. The usual causes are a failed ONNX or policy-manifest load, a network or CDN problem, or a stale browser cache. Reload the page first; if it persists, work through What should I do if the site has issues? under Getting Started.
That's the point: the task's trained policy is executing. Your job is to watch closely and take over only at the moment it's about to fail. If the policy completes the task without your help, that's the best possible outcome — it verifies the policy works, and it scores highest.
Take over at the moment the policy looks like it's about to fail — a bad approach, a slipping grasp, the wrong object — and use the minimum time needed to get it back on track. Don't grab control at the first wobble: intervening early spends budget on states the policy could have handled, and it lowers your score. Don't sit idle while in takeover either; the clock and the trajectory both keep running. Correct minimally, then hand back with Return to Policy.
Each post-training session caps total manual control at 30 seconds, counted across all takeovers in the run. If the budget runs out before the task is complete, the run is automatically marked as failed. Failed runs are not uploaded and don't count against you.
04Points, Scores & Rewards12
Points are Arctivar's non-transferable record of qualifying contribution. Arctivar plans to use Points as one input in a future Community Contributor Airdrop, but Points do not themselves guarantee eligibility, an allocation, or a distribution. Final terms will be published separately. See Points.
Tasks built jointly with ecosystem partners (such as BitRobot), hosted in the Alliance section. Every Alliance task awards Arctivar Points plus a partner reward; eligibility rules and availability are shown in the Hub and governed by the applicable campaign terms.
Yes. Every task carries a difficulty rating from 1 to 5 stars, and the harder it is the more it is worth — with how well you complete it scaling that further. Difficulty and score are two of the five factors below. Hard tasks are where policies still fail, which is why one of them is worth several easy ones.
Not by task count alone. Five factors move your total:
- Volume — how many valid tasks you complete
- Score — the quality of each signed submission; a low-scoring run drags your average down noticeably
- Difficulty — how hard the tasks you take on are, rated 1 to 5 stars
- Diversity — how widely your tasks spread across scenarios and skills
- Referrals — what the contributors you invite actually produce, not how many you invite
The first four are not a checklist and none of them substitutes for another: a hundred trivial submissions in one scenario are worth less than a small set of hard, varied, well-executed ones. See Points.
The average score across your completions of that task is what counts — not your best one. A few sloppy repeats can pull the average down far enough to be worth less than stopping at one good run. Repeat a task to do it better, not to do it more.
For now the two are comparable. Post-training corrections are the scarcer and higher-value data, so they are expected to be worth more over time — but nothing about that is fixed yet, and any change will be announced before it takes effect.
By algorithms, not people — replay verification and sanity checks, rule-based success checking, and a VLM-based motion quality assessment combine into one quality score. No other contributor reviews or grades your runs. For pre-training runs, the score reflects success, efficiency, and motion smoothness; for post-training runs, see What affects my post-training score? under Points, Scores & Rewards. See Points.
Two things — how long the whole run takes, and how much of it the policy drove.
- Trajectory length. Longer trajectories get a lower base score. At 160 samples (32 seconds) the base score reaches 0.
Nis the number of control commands, sampled every 0.2 seconds. - Policy control ratio. The more of the run the policy handled and the less you took over, the higher the score.
Base Score = clamp(1.6 − N / 100, 0, 1) × 100
Final Score = Base Score × (Policy Control Ratio) ^ 0.75The one hard limit is the 30-second takeover budget. It counts only the time you spend in manual control, and running it out fails the run outright — no score, nothing uploaded. Trajectory length is a different quantity: it covers the entire run, policy time included, and it lowers your score rather than failing you.
Points settle on a two-week Epoch. Each Epoch closes on a Friday at 00:00 UTC — that timestamp bounds the calculation — and the Points land in your account on the following Monday at 12:00 UTC. They accumulate Epoch over Epoch; nothing resets or expires. See Points.
Not until you sign them. Only runs signed on-chain, with a transaction recorded on Base, enter an Epoch's Points statistics. A run you completed but left unsigned earns nothing for that Epoch — sign it later and it counts in the Epoch you signed it in. Nothing is lost by signing late. See Only Signed Submissions Count.
If you signed them, yes. When the Point System went live, a one-time retroactive snapshot was taken at 16:00 UTC on 29 July 2026 to credit everything that came before it. Every run you had signed on-chain by that moment was tallied and paid out in a single retroactive grant — nothing needs to be redone, and that phase is now closed.
Work you completed but never signed was not recorded and did not enter the grant. Everything signed after the snapshot falls under ordinary Epoch emission instead. See How Points Are Released.
Because scoring and verification are independent. Signable means the trajectory passed physics validation and the task checker. Score measures efficiency — trajectory length and policy control ratio.
So a run can genuinely complete the task and still score 0: it exceeded 160 samples and the base score fell to 0, or its final score came in below 1 and was rounded down to 0 on-chain.