The rest of this piece is what that means. Not a rule engine, not a set of detectors: a world model, trained the way world models are trained. What that architecture buys you shows up in what the small versions can do.
Why this matters
The reader already lives in a world of driver-assistance features: lane keeping, adaptive cruise, forward-collision warning. Those are ADAS. They are engineered per rule and per sensor, and they perform well on the situations that get labeled in a spec sheet.
A world model is different. It learns to expect what the road usually looks like, and it flags a moment that departs from that expectation. What lets it work on driving in general is scale: not a scale of rules, but a scale of measured driving to learn from.
Empiric Earth trains BADAS on 10B miles of driving, with an asterisk noting the consolidated figure is still being finalized. The training draws on a library of more than 70 million long-tail driving videos.
What a world model is
Concretely: a model that learns to fill in what should come next. BADAS is built on a V-JEPA style architecture that learns in latent space rather than by predicting pixels. Sections of an encoded video vector are masked, and the model is trained to complete them.
That distinction sounds academic. It is not. Predicting pixels is expensive and forces the model to reproduce the appearance of the world. Predicting in latent space forces it to learn the structure instead. The difference shows up later, in how the small models perform.
The most recent shipped version of BADAS was trained on 2 million clips extracted from 200,000 videos. That figure of 200,000 is the currently published figure for the training corpus.
Input is a 2-second video clip. Two seconds is the length the model sees. It is not the warning window: those are different questions.
How it works
One trained world model can drive three different jobs: collision risk, video search embeddings, and trajectory prediction for autonomous vehicles. Most of the model's understanding comes from the self-supervised pre-training stage. Distillation onto downstream tasks mainly tunes where the model draws the line, rather than teaching it what a road looks like.
The most recent shipped version of BADAS 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.
The Flashlight is the model that gets to run on cheap edge hardware. That is the size that matters for a fleet camera in the field.
Where we fit
Public AV research trains on datasets recorded by expert drivers in heavily instrumented cars during uneventful driving. That is not what Empiric Earth's data looks like. Real-world driving carries the edge cases you cannot get on an expert-driven test route. That difference in source drives the difference in what a downstream model can learn.
The training-data pipeline uses the previous BADAS version as an oracle: score every video in the archive, send the suspect ones to human annotators, and confirm the collisions. Positive risk examples are the scarce ingredient. Negative examples are plentiful and repetitive, and the model does not need every one.
Only we know
BADAS worked on new customer data without fine-tuning, repeatedly, against the team's own expectation. Different customers, no adjustment needed to get it to run. The internal name for the effect is "the BADAS moment": a model that carries what it learned in training over to data it has never seen.
The model shows what caught its eye. Heat maps and bounding boxes come straight from BADAS's own attention, on the forward pass, with no extra computation. That visual explainability was an emergent property: BADAS was only ever trained on binary labels, and the attention maps were never a training target. Emergent is a stronger word than trained, and emergent is what the evidence supports.
Limits
BADAS predicts risk. It does not steer. That distinction is deliberate and worth stating on any page that describes what the model does.
The training corpus is chosen carefully rather than sampled uniformly. Positive risk examples are the scarce ingredient, and negative examples are plentiful and repetitive. The 10-billion-mile library does not translate into a 10-billion-mile training set for exactly that reason.
Written natural-language explanations of what happened come from a separate reasoning head that is too heavy to run at the edge, so the attention heat maps are the real-time story and the written reasoning is offline.
FAQ
Q: What is BADAS? A: BADAS stands for Beyond ADAS: a model that reads road video and flags the moment before something goes wrong, built to run on the camera itself.
Q: What is a world model? A: A model that learns what a scene usually looks like and predicts what comes next. BADAS is built on a V-JEPA style architecture that predicts in latent space rather than in pixels.
Q: How much data is BADAS trained 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 version of BADAS specifically used 2 million clips from 200,000 videos.
Q: How large is the model? A: The most recent shipped version of BADAS ships in three sizes: 300 million, 86 million, and 22 million parameters.
Q: How does BADAS explain what it saw? A: Attention maps drawn straight from the model's forward pass, an emergent property of training on binary labels alone.