ROBOTICS PLAYGROUND
Dexterity Lab.
A little pressure. A better result.
Teach a bot a more delicate touch.
Swipe to see all five tasks →
Lift the object and carry it to the pad without dropping or crushing it.
Teach a finer touch
Try 81 settings. Keep the one that works best.
Preparing the first calibration…
Learning…
Each setting runs four seeded trials. The top five get eight validation trials each.
The bot sees exact simulator conditions. It learns two fixed settings per task, using synthetic outcomes.
Same conditions. Paired seeds.
20 evaluation episodes per controller, separate from calibration and validation.
| Controller | Complete | Mean quality |
|---|
Inside this attempt
Live readings from the synthetic environment.
Activity
- Choose a task, then run the bot or take control.
What is actually learning?
This is a small calibration experiment, using a table of sampled rewards. It tests every pair of nine possible values for two controls, estimates quality from four trials each, then validates the top five settings. The chosen settings stay fixed during an automated attempt. The task’s stage sequence and scoring rules are authored.
Changing a condition discards the previous calibration. This is local relearning for one synthetic task, with no neural network or camera perception. Cooking, juicing, force, joints and hair are simplified toy models. The numerical conditions come directly from the simulator. The benchmark measures repeatability under small seeded variation, not performance on unseen materials or real robots.
All simulations run in your browser. No account, external AI service or physical device is connected.