That distinction is the point of this piece. What makes Empiric Earth's data different is not the raw scale of the miles. It is the shape of the events those miles include, and the pipeline built to find them.
Why it matters
A world model for driving learns from what it sees. If it only sees expert drivers doing the normal thing on an instrumented test route, it learns normal driving. Public AV datasets are recorded that way by design: expert drivers on instrumented rigs, avoiding crashes on purpose. Empiric Earth's data is recorded differently, from real people on real roads, and reconstructing the edge cases inside that data lets a downstream model learn from the future trajectories of the third parties involved.
Which brings the argument to the core problem: real-world events are rare. Almost every mile of driving is uneventful, and uneventful data is not what a risk model needs to learn from.
What the training set actually is
Positive risk examples are the scarce ingredient. Negatives are plentiful and repetitive, and a model does not need every one of them. The team collects what it can of the positives, and expects to keep scaling that specifically.
The most recent shipped version of BADAS was trained on 2 million clips extracted from 200,000 videos. The 200,000 figure is the currently published corpus size for the training video count.
Empiric Earth's combined library publishes as 10B miles, with an asterisk noting the consolidated figure is still being finalized. The library holds more than 70 million long-tail driving videos to draw from.
The training-set-to-library ratio is intentional. Two million clips against more than 70 million videos, and the whole system is built to find the interesting minority.
How the data pipeline works
The mining strategy runs the previous version of BADAS as an oracle. Score every video in the archive with the prior model, send the suspect clips to human annotators, and confirm the collisions. The model helps find its own next training set.
That is a genuine flywheel, and it explains why version-over-version data expansion has held. The most recent shipped BADAS trained on five times the data of the version before.
Most of the model's understanding comes from the self-supervised pre-training stage. Distillation onto the downstream task mainly tunes where the model draws the line rather than teaching it what a road looks like.
The architecture that lets this work is a V-JEPA style world model: sections of an encoded video vector are masked, and the model predicts what should complete them, learning in latent space rather than in pixels.
Where Empiric Earth fits
One trained world model can drive three different jobs: collision risk, video search embeddings, and trajectory prediction. The same base is what Empiric Earth ships in three sizes: a 300-million-parameter Large, an 86-million-parameter Flash, and a 22-million-parameter Flashlight. The Flash and Flashlight versions are built by distilling the Large model.
Small models are the ones that get to run on edge hardware. They inherit their understanding of driving from the pre-training that was done at scale, and the distillation step is what fits that understanding into a footprint that can sit on a camera.
Only we know
BADAS worked on new customer data without fine-tuning, repeatedly, and across different customers. That was against the team's own expectation going in. The generalization is the tell that the model is picking up structure of a road rather than the appearance of the specific customer's data.
Limits
The 10-billion-mile library does not translate directly into a 10-billion-mile training set. It cannot. Positives are the scarce ingredient, and pretending otherwise is where the "we have more miles than everyone" claim breaks.
FAQ
Q: How much data does BADAS train on? A: Empiric Earth's combined library publishes as 10B miles, with an asterisk noting the consolidated figure is still being finalized. The most recent shipped BADAS was trained on 2 million clips extracted from 200,000 videos.
Q: Why isn't the training set larger? A: Positives (real risk events) are the scarce ingredient. Negatives are plentiful and repetitive, and a model does not need all of them.
Q: How does Empiric Earth find the positives in a huge archive? A: The previous version of BADAS scores every video, and the suspect clips go to human annotators for confirmation.
Q: What is the architecture that learns from this data? A: A V-JEPA style world model that predicts in latent space rather than in pixels.
Q: What is the training set for the smaller models? A: The Flash and Flashlight versions are distilled from the Large model, so they inherit the pre-training that was done at scale.