Post-training Tasks
Watch the robot try the task, then take over to fix it — how post-training sessions work.
In a pre-training task, you show the robot how to do a task from scratch. In a post-training task, the robot already knows — imperfectly. A policy trained on contributor demonstrations attempts the task on its own, and your job is to catch it the moment it slips, correct the mistake, and hand control back.
Your corrections are the highest-value data on the platform: they capture exactly the situations the policy cannot yet handle, and they train its next iteration. The stronger policy then reaches new situations it cannot handle — and those define the next round of post-training tasks. For the theory behind this loop, see Model Training.
The Session
A post-training task looks like a pre-training task — same goal, steps, controls, and simulator — plus a Run Control panel on the right.

Run the policy
Press Run Policy. The robot begins attempting the task autonomously while you observe. A status banner reminds you: the policy is inferring — watch closely, and take over the moment it might fail.

The policy drives; the button becomes Take Over. Take over
The moment the policy looks like it is about to fail — a bad approach, a slipping grasp, a wrong object — hit Take Over to grab manual control. Deciding when to intervene is the skill: too early wastes your takeover budget on states the policy could have handled; too late lets the failure happen.
Correct it
Now you're driving, with the full pre-training control scheme: keyboard, hold and drag the blue ring on the arm, or double-click an object to send the gripper to it. Fix what the policy got wrong — reposition, re-grasp, complete the step. Undo or Redo a slip, or Reset to start the run over.

Takeover active: correct the task, then return control. Return to policy
Once the robot is back on track, press Return to Policy. The policy resumes execution and continues the task. You can take over again if it slips later — the session ends when the task is completed (or failed).
The Takeover Budget
Each session has a takeover budget of 30 seconds — the total time you can spend in manual control, across all takeovers in the run. The counter runs the whole time you are in control.
The budget is not a formality — it is what makes the data valuable. A correction dataset is useful precisely because the policy does most of the work and humans supply only what it lacks. Long takeovers turn a correction back into a demonstration, which is what pre-training tasks are for.
How Sessions Are Scored
The task must be completed. This is the hard rule, and the 30-second takeover budget above is what enforces it: incomplete runs, budget overruns included, earn no score and no Points.
Within a completed run, two quantities set the score:
- Trajectory length. The longer the whole run takes — policy time included, not just yours — the lower the base score. At 160 samples, meaning 32 seconds at one control command every 0.2 seconds, the base score reaches 0.
- Policy control ratio. The larger the share of the run the policy drove, the higher the score. Less takeover is better.
Base Score = clamp(1.6 − N / 100, 0, 1) × 100
Final Score = Base Score × (Policy Control Ratio) ^ 0.75N is the number of control commands in the trajectory.
A run where the policy completes the task with zero intervention is the best possible outcome — it verifies the policy works in that scenario. A run with one short, precise correction at the exact failure point is the next best. Watch the Policy success rate on the task panel to gauge how often the current policy makes it on its own.
Reproducible by Construction
The policy runs directly in your browser — the same place the physics runs — so there is no round-trip to a server between observation and action. Every rollout is seeded: the initial scene state and the policy's execution are reproducible, so your correction can be replayed and verified exactly as recorded. This is also why corrections are so effective as training data — each one is anchored to a precise, recoverable failure state.
Slots and completion limits work like every other task: each session consumes one slot, and you can complete a given task up to 5 times — with the average of those completions, not the best, being what the task is worth to you.
Tips for Better Corrections
Next: Sign & Verify →