✳ ROBOT LABEXPLORE · LEARN · CHALLENGE
Your new robotBrowser preview
Can a skill transfer to a new maze?
Untrained browser preview
0 moves · 0 points
ROBOT LAB

Run details

Your robot's run

Not training
  1. Queued
  2. Preparing
  3. Training
  4. Testing
  5. Ready

The browser tests a model; the server GPU trains it.

Learning time—
Parallel worlds—
Practice moves—
Recorded training examples & timeline

Saved batch examples or states from this run. These are not extra robots training in your browser.

What happens on the GPU?

The unfamiliar-maze mode learns spatial weights from varied teacher examples. A recurrent neural planning layer applies those learned weights across each map. The four movement directions are built into that structure; routes are not stored.

Training examples use a pathfinder as a teacher. During tests, the neural network chooses the moves without calling that teacher. Practice checks can guide difficulty; the final exam is separate and never supplies training gradients.

Runs use the existing shared student queue and a capped turn. A small model may use little of the GPU's memory. Budget labels describe generated tasks, not guaranteed unique layouts or guaranteed success.

Keep your robot

Choose an unused name. This browser remembers your robot; download your work before clearing browser data.

Two ways to learn

Already saved a robot at the old address? Open the MRC projects page in the same browser and choose Open Robot Lab to bring your saved robot here.

Learn my own maze: drive a route, then train using your examples and rewards. Compare results to spot memorisation.

Learn unfamiliar mazes: a grid-based model learns from changing teacher examples. It begins with short routes, adds obstacles and then corridors when practice checks reach 80%. It finishes with a separate, frozen 100-maze exam.

  1. Predict: design a challenge and write what you expect.
  2. Observe: try the starting model and inspect its decisions.
  3. Train: choose a budget and join the shared GPU queue.
  4. Challenge: compare models on a new maze without changing their weights.
  5. Explain: use the measured results and download your work.

The unfamiliar mode uses a learned spatial cost network and a built-in recurrent planning structure. Teacher paths are used only to train it. It does not call a pathfinder during browser playback or exams. This structure helps transfer what it learns to new maps; success is still measured, not assumed.

Practice validation guides learning difficulty; it is not the final exam. The exam uses different seeds and never trains the model. Quick budgets may remain on easy stages.

Older models and earlier versions remain saved. Choose the same starting model when switching learning modes. Existing v1 models use 9×9 boards.

Click any statistic to expand it. Live updates preserve open panels, notes and the map view. On a small screen, Lesson controls opens the sidebar over the map.

Saved model versions

Restoring a version preserves the current model.