The Driving Model is the AI brain behind every steering, gas, and braking decision your car makes while FrogPilot is engaged. FrogPilot lets you swap between multiple community-contributed driving models, each trained slightly differently, so you can find the one that drives best on your roads. You can download new models, rate them after each drive, and even let a randomizer cycle through them to help you discover your favorite.
What It Is
The driving model is a neural network that processes live camera footage and turns it into a complete driving plan: where to steer, when to brake, and how fast to go.
Think of it like choosing a driving instructor. Every instructor follows the same rules of the road, but each has their own style: one might take curves more smoothly, another might brake a bit earlier at stop signs, and a third might handle highway merges more confidently. FrogPilot's driving models work the same way. They all use camera input to produce steering and speed decisions, but their different training data and architectures mean they behave slightly differently in real-world situations.
Each model is made up of two core components. The vision encoder processes raw camera frames to understand what's in front of you: lane lines, other cars, traffic signals, road edges, and curves. The driving policy takes that understanding and decides what the car should actually do: the exact steering angle, the target speed, and the acceleration or braking needed.
FrogPilot hosts additional models in an online repository beyond what stock openpilot offers. These models are downloaded to your device and can be swapped without reinstalling anything. The active model's name is displayed on your home screen next to the FrogPilot version number, so you always know which model is driving.
Before You Start
The Driving Model panel is accessible at Standard tuning level or higher.
If your tuning level is set to Minimal, you won't see the "DRIVING MODEL" button inside Driving Controls. Switch to Standard, Advanced, or Developer to access model selection and management.
The built-in default model, Dark Souls, is always available without any downloads. To try other models, you'll need:
- An internet connection (Wi-Fi or cellular) to download models from the FrogPilot model repository
- Adequate storage on your device, since each model consists of multiple files
- Your car to be parked (the device won't start a download while driving)
⚠️ Warning: Switching models while the car is running requires a reboot to take effect. FrogPilot will prompt you to reboot when you select a different model with the car on.
Settings
Every setting in the Driving Model panel controls which model runs, how new models arrive on your device, and how you evaluate them.
Settings Summary
| Setting | Default | Tuning Level | What it does |
|---|---|---|---|
| Select Driving Model | Dark Souls (Default) 👀📡 | Standard | Choose which model drives your car |
| Automatically Download New Models | On | Standard | Keeps your model library up to date |
| Download Driving Models | — | Standard | Manually fetch specific models |
| Delete Driving Models | — | Standard | Free storage by removing models you don't use |
| Model Randomizer | Off | Advanced | Picks a random model each drive for comparison |
| Manage Model Blacklist | — | Advanced | Exclude models from the random rotation |
| Manage Model Ratings | — | Advanced | View or reset your per-model rating history |
| Update Model Manager | — | Standard | Updates the internal engine that runs newer models |
Select Driving Model
| Property | Value |
|---|---|
| Default | Dark Souls (Default) 👀📡 |
| Recommended | The default Dark Souls model works well for most drivers |
This is where you pick the model that FrogPilot uses to drive. When you tap SELECT, you'll see a list of all downloaded models with emoji indicators showing each model's capabilities:
| Emoji | Meaning |
|---|---|
| 🗺️ | Supports navigation-aware driving |
| 📡 | Supports radar input |
| 👀 | Supports Vision Only Adaptive Cruise Control (camera-only gas and braking) |
The default model, Dark Souls (Default) 👀📡, supports both camera-only cruise control and radar input. It's always present on your device and can't be deleted.
What this means for you: If you've never changed your model, you're using Dark Souls, and your selection will automatically follow the recommended default when FrogPilot updates. Once you explicitly select a different model, your specific choice is preserved across updates.
💡 Tip: When Model Randomizer is enabled, the Select Driving Model option is hidden since the system picks for you automatically.
Automatically Download New Models
| Property | Value |
|---|---|
| Default | On |
| Recommended | On |
When On, FrogPilot periodically checks the model repository for new or updated models and downloads them in the background. It compares the file sizes of your local models against the remote versions and re-downloads any that have changed.
When Off, you'll need to manually download models using the Download Driving Models button. Already-downloaded models continue working normally.
What this means for you: Leave this on to always have the latest models available. Turn it off if you have limited storage or a slow internet connection and prefer to manage downloads yourself.
Download Driving Models
This button lets you manually download individual models or all available models at once:
- DOWNLOAD: Opens a list of models that aren't on your device yet. Pick one to start downloading.
- DOWNLOAD ALL: Downloads every available model that you don't already have.
During a download, the button changes to CANCEL so you can stop it at any time. A progress percentage shows how far along the download is. FrogPilot tries GitHub first, then falls back to GitLab if GitHub is unavailable.
⚠️ Warning: Downloading must happen while parked and connected to the internet. If a download fails or is cancelled, partially downloaded files are cleaned up automatically.
Delete Driving Models
| Property | Value |
|---|---|
| Default | — |
| Recommended | Use when storage is low |
Remove downloaded models you no longer want. You can delete them individually or all at once. Two models are protected and can't be deleted: the currently active model and the default model (Dark Souls).
Model Randomizer
| Property | Value |
|---|---|
| Default | Off |
| Recommended | On if you want to find your best model through blind comparison |
When On, FrogPilot selects a random model from your downloaded collection each time you finish a drive and park. The model name is hidden and shown as "Mystery Model 👻" so your expectations don't influence your judgment. After drives lasting more than 15 minutes, you'll see a rating screen where you can score how the model performed.
When Off, FrogPilot uses whichever model you've explicitly selected.
What this means for you: This is like a blind taste test for driving models. By hiding which model is active, you rate purely based on how the drive felt, not on name recognition. Over time, the rating system surfaces which model genuinely drives best for your routes.
| Scenario | Randomizer Off | Randomizer On |
|---|---|---|
| Model used | Your explicit selection | Random from downloaded pool |
| Blacklisted models | Not applicable | Excluded from pool |
| Model name shown | Actual name | "Mystery Model 👻" |
| Rating prompt after drive | Drive Summary only | Rating screen (drives over 15 min) |
| Consistency | Same model every drive | Different model each drive |
When you first enable Model Randomizer, FrogPilot will ask if you'd like to download all available models so the random pool is as large as possible.
Manage Model Blacklist
| Property | Value |
|---|---|
| Default | Empty (no blacklisted models) |
| Recommended | Blacklist models that consistently score poorly |
This setting only appears when Model Randomizer is enabled. It gives you three options:
- ADD: Pick a model to exclude from the random rotation
- REMOVE: Bring a blacklisted model back into the rotation
- REMOVE ALL: Clear the entire blacklist
You must always have at least one model that isn't blacklisted. If you try to blacklist every model, FrogPilot will prevent it.
What this means for you: After a few weeks of randomized driving, you'll likely have a few models that consistently score low. Blacklisting them shrinks the pool so you only rotate through models that drive well on your roads.
Manage Model Ratings
| Property | Value |
|---|---|
| Default | No ratings stored |
| Recommended | Review periodically to find your top model |
This setting only appears when Model Randomizer is enabled. It provides two options:
- VIEW: Opens a screen showing every available model along with its score percentage and drive count
- RESET: Clears all rating data and drive counts so you can start fresh
Each model's data includes:
| Field | What it means |
|---|---|
| Score | A running percentage (0–100%) based on your emoji ratings |
| Drives | How many drives you've completed with that model |
The score is a weighted running average. Each new rating adjusts the overall score based on how many drives you've accumulated. A model with 20 drives and an 85% score is a more reliable indicator than one with 2 drives and 100%.
⚠️ Warning: Rating data is intentionally excluded from toggle backups. If you factory reset your device, all model ratings and drive counts will be lost.
Update Model Manager
| Property | Value |
|---|---|
| Default | — |
| Recommended | Update whenever prompted |
The Model Manager is the internal engine that loads and runs newer-generation driving models. When an update is available, this button shows "Update available!" and lets you download the latest version.
Updating the Model Manager will delete all existing newer-generation model files, which will then need to be re-downloaded. FrogPilot handles this automatically: after the update, it re-downloads your currently selected model first, then any other models you had installed.
When the Model Manager is already current, the button shows "Up to date!" and is disabled.
How It Works
FrogPilot manages a complete model lifecycle: discovering available models, downloading them, selecting one to run, and collecting your feedback.
Discovery and Downloads
FrogPilot periodically checks the online model repository for a list of available models. This list includes each model's name and version number. When new models appear or existing ones are updated, FrogPilot can automatically download them if you have Automatically Download New Models enabled.
Downloads happen from GitHub as the primary source, with GitLab as a fallback. Each download is verified by comparing the downloaded file size against the expected size on the server. If verification fails from GitHub, FrogPilot automatically retries from GitLab.
Older models (versions 1 through 6) consist of a single file per model. Newer models (version 7 and above, including the default Dark Souls v9) consist of four files per model: the vision model and its metadata (which process what the camera sees), plus the driving policy model and its metadata (which decide how the car should drive based on what the camera saw).
Model Selection and Loading
When you select a model (or when the randomizer picks one), FrogPilot stores your choice. The model is loaded fresh each time the car starts driving. If the car is already running when you change models, you'll need to reboot because the device's process manager only loads the model at the start of a drive.
If Model Randomizer is enabled, the random selection happens when you finish a drive and park. FrogPilot filters out any blacklisted models and any models that aren't downloaded, then picks randomly from the remaining pool. If the pool is empty (all models blacklisted or none downloaded), it falls back to the default Dark Souls model. Your next drive then uses this freshly randomized selection.
The Default Model Behavior
If you've never manually selected a model, your device uses the current recommended default (Dark Souls v9). This means if FrogPilot updates and the recommended default changes in the future, your device automatically picks up the new one.
Once you explicitly select a specific model through the settings, your choice is locked to that exact model and won't change with updates, even if the recommended default shifts.
Rating and Scoring
After any drive where Model Randomizer is enabled and the drive lasted more than 15 minutes, FrogPilot shows a rating screen instead of the normal drive summary. You'll see five emoji buttons representing your satisfaction:
| Rating | Emoji | Score Value |
|---|---|---|
| Excellent | 🤩 | 100 |
| Good | 🙂 | 80 |
| Okay | 🤔 | 60 |
| Poor | 🙁 | 40 |
| Terrible | 🤕 | 20 |
The score formula is a weighted running average:
In plain English: Each new rating blends with all your previous ratings for that model. Early ratings have a big impact, but as you accumulate more drives, individual ratings smooth out. A model's score stabilizes over time, giving you a reliable picture of how well it performs.
After rating, you'll see the model's current rank (compared to all rated models), its score percentage, the number of drives with that model, and your total drives across all models. You can also blacklist the model directly from this screen if the drive was particularly bad.
Model Generations
Different model versions run through different processing engines on your device. FrogPilot automatically detects the model's version and starts the correct engine each time the car begins a drive:
| Model Version | Description |
|---|---|
| v1 – v4 | Older "classic" models, single file each |
| v5 – v6 | Intermediate models, single file each |
| v7+ (including Dark Souls v9) | Latest generation, four files each, using the newest processing engine |
You don't need to worry about which engine runs. FrogPilot handles this automatically based on the model you select.
Limitations & Known Issues
No single model is universally best. Performance varies by car, road type, weather, and lighting.
- Model performance is subjective. A model that handles highway curves beautifully might struggle in tight urban intersections. Your ideal model depends on your typical driving conditions.
- Rating data doesn't survive factory resets. Scores and drive counts are intentionally excluded from toggle backups, so all rating history is lost if you reset your device.
- Downloads require connectivity and storage. Each newer-generation model consists of four files. Downloading all available models requires both a stable internet connection and sufficient free space on your device.
- Model changes require a reboot while the car is on. If you're parked with the ignition off, the new model loads automatically on your next drive. If the car is already running, you'll need to reboot because the device's process manager only loads a model when a drive begins.
- Model Randomizer hides the model name intentionally. You won't know which model drove until the rating screen appears (drives over 15 minutes) or until you check the home screen after the drive.
- The model version determines which internal processing engine runs. Older models (v1 through v6) use different engines than newer models (v7+). This is handled automatically but means that switching between very old and very new models involves a different software pipeline.
- Neural Network Feedforward is independent of model choice. Neural Network Feedforward availability depends on whether your specific vehicle has a matching steering model file, not on which driving model you select. Changing your driving model won't enable or disable Neural Network Feedforward.
Setup Recommendations
Start with the default and only explore alternatives if you're curious about differences.
| Scenario | Recommended Setup |
|---|---|
| New to FrogPilot | Leave everything at defaults. Dark Souls (v9) is well-tested. |
| Curious about alternatives | Enable Automatically Download New Models, then try different models one at a time via Select Driving Model. |
| Systematic comparison | Enable Model Randomizer, download all models, drive for a few weeks, then check Manage Model Ratings to see which scored highest. |
| Low storage device | Disable Automatically Download New Models and only download specific models you want to try. Delete ones you're done with. |
| Found your favorite | Disable Model Randomizer, select your top-rated model, and enjoy consistent drives. |
Q: I don't see the "DRIVING MODEL" button in Driving Controls. Where is it?
A: The Driving Model button only appears when your tuning level is set to Standard or higher. If you're on Minimal, navigate to FrogPilot settings and change your tuning level to Standard.
Q: My selected model disappeared after an update.
A: FrogPilot checks your selected model against the current list of available models. If your model was removed from the repository or renamed, FrogPilot resets to the default Dark Souls model. Enable Automatically Download New Models to keep your library current, or manually re-download the model if it's still available.
Q: Model Randomizer keeps picking the same model over and over.
A: Check how many models are actually in the random pool. If you've only downloaded two models and blacklisted one, the randomizer has only one option. Download more models or remove entries from your blacklist. Remember that the randomizer only picks from models that are both downloaded and not blacklisted.
Q: A download seems stuck or keeps failing.
A: First, confirm your device is connected to the internet and parked (downloads are disabled while driving). Check that you have enough free storage on your device. If GitHub is experiencing downtime, FrogPilot will automatically try GitLab as a fallback. If both fail, you'll see a "GitHub and GitLab are offline" message. Wait and try again later.
Q: My model ratings and scores disappeared after a factory reset.
A: This is expected. Rating data is intentionally excluded from toggle backups, so scores and drive counts don't persist across device resets. You'll need to start your model evaluation from scratch.
Q: How do I know which model is currently active?
A: The active model's name is shown on the FrogPilot home screen, right next to the software version number in the top-right corner. If Model Randomizer is enabled, it will show "Mystery Model 👻" while the model name is hidden, but after the drive ends, you can see which model was used on the rating screen.
Q: The "Update Model Manager" button says "Update available!" but I'm not sure what it does.
A: The Model Manager is the software component that loads and runs newer-generation driving models. When an update is available, it means there's a newer version of this component. Tap UPDATE to download it. Be aware that updating will delete all existing newer-generation model files, which FrogPilot will then re-download automatically. The currently selected model is re-downloaded first so it's ready for your next drive.
Q: I blacklisted a model by accident from the post-drive rating screen. How do I undo it?
A: Go to FrogPilot → Driving Controls → Driving Model → Manage Model Blacklist, then tap REMOVE. You'll see a list of all blacklisted models. Select the one you want to restore to the random rotation.
Q: What's the difference between the "Driving Model" panel and "Model UI" in Theme and Appearance?
A: The Driving Model panel (in Driving Controls) controls which AI model makes driving decisions. Model UI (in Theme and Appearance → Appearance) controls how the model's predictions look on screen: the width of the driving path, lane lines, road edges, and path edges. Changing Model UI settings has no effect on how the car drives, only on what you see.
Q: Can I use an older model version with newer FrogPilot features?
A: Yes, FrogPilot automatically detects the model version and loads the appropriate processing engine. Older models (v1 through v6) use a different engine than newer ones (v7+), but FrogPilot features that depend on model output (like Conditional Experimental Mode and Curve Speed Controller) work with any model. However, newer models generally produce better driving predictions.
Q: The score formula seems simple. Does a single bad drive ruin a model's rating?
A: Early on, yes, individual ratings have a significant impact. But as you accumulate more drives, each new rating has less influence. For example, if a model has a 90% score after 10 drives, one "Terrible" (20) rating would only drop it to about 84%. After 50 drives, the same bad rating would barely move the needle. The system is designed to stabilize over time.
