Contributor Guide

Pre-training Tasks

What happens from the moment you enter a pre-training task to when your demonstration is uploaded.

In a pre-training task, you teleoperate the robot to complete the task from scratch. One clean run through the steps is a demonstration — the raw material that trains the task's base policy. (For tasks where a trained policy already exists, see Post-training Tasks.)

The Task Interface

When you enter a task, the screen switches to a full-screen view. The simulated environment renders in real time using MuJoCo (WebAssembly) and Three.js — all running locally in your browser.

Pre-training task interface: goal and controls on the left, live simulator with arm cameras in the center, slots, pass rate, and completion-time distribution on the right
AreaWhat it shows
Top barTask name, timer, latency and FPS
Left panelGoal, numbered steps, and the full control reference (pinned — scroll it to see everything)
CenterThe live simulator, with arm camera picture-in-picture views and a sensitivity slider
Right panelRemaining slots, difficulty, pass rate, the completion-time distribution and top ranking, and a goal reference image of the finished state

The goal reference shows what the end state should look like — your layout may vary slightly, since initial conditions are sampled per run.


Controlling the Robot

You control the robot's end effector in task space — you move the gripper's position and orientation, not individual joints. An inverse kinematics solver resolves your target into joint-level commands, and execution follows smooth online interpolation in joint space, keeping motion steady even on bimanual embodiments.

Three input methods, freely mixed:

  • Drag — hold and drag the blue ring on the arm to move the gripper along the current view plane.
  • Double-click — double-click an object and the arm moves into a pre-grasp pose on its own, skipping the uninformative travel across free space.
  • 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. See the full control reference.

Concentrate your effort on the parts of the trajectory that carry the manipulation itself: grasp selection, and the fine adjustments immediately before and after gripper closure.

Real-Time Feedback

As you work through the task, the step tracker follows your progress. Each step lights up when the system detects completion — for example, "Object picked up" → "Object moved to target zone" → "Object placed successfully."

Behind the scenes, every control tick logs the full simulator state:

  • q_robot[t]: robot joint states and velocities
  • action[t]: your control commands
  • The initial scene state (object poses and physics state), recorded as a seeded snapshot

Because the run is seeded and the complete state is captured, the entire scene — including every object's motion — can be re-simulated deterministically. That is what makes replay verification possible. This data forms your trajectory.


Completing a Task

The system automatically detects when success conditions are met — typically, the target object reaching the goal region without violating constraints (e.g., no catastrophic collisions, objects intact).

On success, a Task Completed popup displays your completion time, and your trajectory is packaged and uploaded automatically. The quality score is not part of the popup: it is assigned only after the run has been verified by replay, and it shows against the run in your Portfolio's History tab (see What Happens to Your Data).

You can immediately start again to record another demonstration. There is no cooldown — each task can be completed up to 5 times per user. Check the distribution panel to see how your time compares with other contributors, or Top Ranking for the fastest runs.

Failing a Task

There are only two failure conditions:

  1. Time limit exceeded — for tasks with a time constraint.
  2. Slots filled — all remaining data slots were claimed while you were playing.

Failed attempts are not uploaded and do not count against you. You can retry immediately (unless slots are full).

Tips for Better Data

Next: Post-training Tasks →
Pre-training Tasks