AI Training Story: Driving Around is a standalone controller workshop inspired by AI Learns To Drive. Train a population, evaluate an unchanged driver, and inspect its measured replay. You or an agent can operate the experiment tools; the opening question is a challenge to investigate through matched comparisons.
Engine forked from Zorro909’s driving_sim_rs; training interface and hosting by TiGa. Inspired by the game by ActivationFunction. Built-in tracks are reconstructed replicas. For supported conditions, the lab checks the run’s final fitness policy (or frozen input) with a bounded fresh replay. It predicts conditions only and does not award or accumulate original-game stars; opponent races and other unsupported conditions remain unchecked.
Uploaded files are used for the requested run. Results are held briefly in application memory and deleted after verified download or expiry. Browser recipes contain settings only. Read the data policy.
The map shows the road and walls from native track-part geometry. An optional suggested route gives a rough reference alongside the car’s recorded path. Record a short CPU drive to watch an unchanged policy using real vehicle positions. A new recording can differ from an earlier training run. Native control commands and consumed sensor inputs appear in the black box. Ray-hit positions and neural activations are not recorded.
Report a problem · Opens an editable draft for you to review.
Build: connecting…
Use the same local controls as the lab: read the current state, open a section, create a starter, or inspect the active driver. These tools do not submit jobs or save data. Agent interface guide · Action manifest (JSON)
Checking browser tool support…
The state below is also available to ordinary browser automation. Each listed action identifies a stable control in this page.
Starting fresh? Choose a vehicle and create a starter neural network to train. Already have a driver? Load its file or bucket bundle.
Built-in replica or your own track.
Short first-run defaults are ready. Adjust them or save a recipe.
Train, download and replay. A first run may not finish a lap.
The received driver and analysis stay in this page until discarded, replaced, or closed. Save downloads to keep them. Reports describe server-reported measurements; reward weights are configured values, not measured contributions.
Continue with a saved driver, or create a new network to start.
Your existing .evolution driver keeps its vehicle, inputs and weights.
Bucket storage is optional. Authorizing storage also identifies your account for GPU requests; ordinary GPU sign-in remains available with the basic identity permission. Then enter your existing private bucket ID. Connected runs automatically save by default; uncheck this option to keep run saves manual. No personal token is requested or stored.
HF describes the requested write-repos permission as read/write access to your personal repositories. That grant is broader than one selected bucket and is not a bucket-only permission; the app limits its own operations to the bucket you enter. The minimum current scope for bucket read/write has not been confirmed by HF.
Checking optional HF bucket authorization…
For each completed run, the browser fetches and verifies the network, then confirms server deletion before sending the network and sanitized log to your bucket. If deletion is not confirmed, no save occurs. If bucket saving is not confirmed, the verified local candidate stays available to download or retry; the remote save may have completed. Each batch step saves before the next starts, and a failed save stops the batch.
Save and import the validated recipes and batch definitions from this browser. This 32 KB library contains settings only; networks, tracks, run results and access tokens are never included. Import asks before replacing the local recipe library and never starts a run.
No recipe library bucket action requested.
Not connected. Bucket storage is optional.
No driver selected
Choose an uploaded file or use a network from the browser editor.
Choose a vehicle below, then create its starter neural network. It starts untrained: choose a track and train it to learn to drive. No existing file is needed.
Creates a small network in this page only. It does not start a run or change saved recipes.
Loading built-in replica tracks…
Training uses the selected recipe. Frozen evaluation runs the selected policy in a fresh car without changing its weights.
Use the current training settings, or choose a saved recipe.
Checking available backends…
Checks server availability. Does not restart training, sign you in, or reset GPU allowance.
CPU needs no sign-in. Checking optional GPU account connection…
CPU works without an account. HF sign-in identifies GPU requests; it does not guarantee available allowance. Sign in before choosing files: sign-in reloads this page and clears selected files and unsaved edits. Saved recipes remain.
Your account token stays in page memory. Refreshing or closing this page disconnects the app. Disconnecting here does not sign you out of the HF website or delete job data.
Checks which account the GPU submission service sees. Runs automatically before GPU training; it does not check your remaining allowance.
Loading selected track…
Calculate an optional route for general orientation, then compare it with the recorded car path.
All values below belong to the selected native sample. Sensor inputs were consumed before physics at the decision tick; vehicle state is captured after physics. Samples are 0.1 seconds apart. Contact events identify intervals, not exact impact ticks; the counter does not count unique crashes.
| Vehicle state | Recorded value |
|---|
| Control | Raw command | Applied after normalization | Source |
|---|
| Input index | Sensor | Parameters | Native input value |
|---|
These are the input values the policy consumed. Ray inputs are not ray-hit positions. Hidden-layer activations are not recorded.
Metrics accumulate at native statistics callbacks; they are not instantaneous forces or per-reward contributions.
| Metric | Recorded cumulative value |
|---|
No complete analysis is available yet.
A network can fill more than one winner role. Select roles to compare; duplicate policies run once. Fitness is final-generation population-relative and is not compared across generations or policy evaluations.
| Compare | Role | Candidate | Generation / car | Selection fitness | Best lap | Forward laps | Physical contacts | Distance (simulation units) | Action |
|---|
Comparison starts one frozen evaluation per distinct selected policy, in sequence. It uses the currently selected track, vehicle, tick limit, control rate and collision/idle settings, then holds those conditions fixed for the full comparison.
Frozen evaluation does not select winners or compare population-relative fitness. Each comparison row keeps its network hash and must match the same track, vehicle, horizon, control rate, collision/idle rules, backend and engine build.
| Policy | Vehicle | Track | Ticks | Best lap | Forward laps | Physical contacts | Distance (simulation units) | Comparison |
|---|
| Challenge | Result | Reason | Completion time |
|---|
This is an independent check of the named run policy. It does not grant or accumulate original-game stars; replica challenges may differ from the original game.
Optional, read-only check for supported built-in tracks. It sends the track ID and best lap to this app server; Steam receives only a public leaderboard read. No score is submitted. The result is an estimate, not an official rank.
Build, export, or restore one validated bundle. Bucket storage is optional. When connected, automatic save is on by default; turn it off here to keep saves manual. Disconnected runs never call bucket storage.
No bucket action has been requested.
Ready. Inputs remain on this device until you run the test.
After a run, supported conditions are checked against its named policy with a bounded fresh replay. Results appear in the Star challenge check panel when available. These are predictions for the reconstructed replica track, not original-game awards; opponent races and unsupported conditions are left unchecked.
A recipe saves training settings and rewards. Apply one to this form, or adjust the settings below and save your own. Choose your track and starting network in Run & results.
Working settings · not saved yet
Recipes and batches are saved in this browser. Export a backup to move them to another browser or site. Clearing site data removes this library.
The recipe library contains settings only. Saving a recipe does not start a run or upload tracks, networks, or results.
Adjusting this form does not change a saved recipe until you save it.
Loading built-in replica tracks…
Keep selected sensor inputs, hidden layers, or output controls unchanged while training. Choices follow the selected network and are saved with recipes. Frozen evaluation always uses an empty freeze selection.
Arrange saved recipes into a sequence of up to 20 training steps. Use the same recipe more than once to keep training with those settings.
New batch · not saved yet
Step 1 uses your selected network. Later steps continue from the previous step’s result.
Save a recipe in Training & recipes, then add it here.
The same steps as the list above. Use saved slots A–J, separated by commas; repeated slots are allowed. Maximum 20 steps.
No batch running.
An original editor inspired by Mars’s network creator. Create fresh networks with the selected vehicle’s game-legal controls, or use its largest original-game shape. Loaded networks keep their original vehicle, output order and sensor parameters. Unsupported files can still be inspected without changing the original.
Inspect the active driver, choose a file, or create a fresh network. Editor changes stay in this page until you use or export them.
| Record | Time | Mutation | Best score | Best lap |
|---|
| Connection / bias | Value |
|---|