Our 3rd open-source driving dataset is live on Hugging Face! 🚗 This time, we’re adding a new layer to NATIX data: vehicle telemetry. Real-world multi-camera footage, now paired with precise signals showing what the vehicle was actually doing.

2/ Why does that matter? World Models and end-to-end driving models need more than video. Wheel angle, pedal position, vehicle speed & automotive-grade GPS let models learn not only what happened on the road, but also the vehicle's precise actions.

3/ But telemetry gives us another powerful capability: finding the long tail. 🔍 We can analyze vehicle signals to automatically identify moments like harsh braking, harsh acceleration, sharp turns & aggressive driving patterns across large amounts of driving data.

4/ Telemetry and VLMs are complementary. Telemetry can flag the exact moment a harsh brake happens. A VLM can then analyze those few seconds of footage to explain why. Vehicle telemetry finds the event. Visual AI explains it. 🤝

5/ The first release includes 5 safety-critical events across 10 U.S. states. Just like our previous datasets, we’re making it openly available for researchers & open-source teams. Read more about what makes this dataset unique 👇

@huggingface The telemetry layer makes this dataset way more useful for training driving models.

@huggingface World models need actions not just pixels, this dataset should really help a lot

@huggingface This gives researchers a lot more to work with.

@huggingface Open datasets like this can seriously accelerate autonomous driving research.

@huggingface The company that got hacked?

@huggingface pairing both could give researchers a much richer dataset for training and evaluating autonomous driving systems.

@huggingface nice addition with the vehicle telemetry.

@huggingface Being able to search for specific moments across huge amounts of data could save researchers a ton of time.

@huggingface Open sourcing safety critical clips with real vehicle data is rare and useful.

@huggingface This is the kind of real-world data that actually helps AI understand driving the way people do, not just how it looks from the cameras.

@huggingface Seeing what the car did matters just as much.

@huggingface Camera feeds see the way, telemetry teaches the real physics.

@huggingface pairing real-world footage with vehicle telemetry gives driving models the context they’ve been missing.

@huggingface how much longer till we can reliably have the dataset & model self-driving? any estimates?

@huggingface With this open-source framework, I bet that in terms of communication, the project has presented many angles for implementation.

@huggingface nice work putting this on huggingface researchers will actually use it

@huggingface can telemetry and VLMs make driving data way better than ever?

@huggingface nice drop. adding actual vehicle telemetry to the multi-cam footage is a big upgrade

@huggingface Open-sourcing the long-tail events might be the right call yunno

@huggingface More context means better training signals for driving models.

@huggingface The telemetry is what actually makes this useful. Video alone is everywhere already. I love the process.

@huggingface Pairing the footage with what the vehicle was actually doing makes this dataset way more useful for training

@huggingface telemetry layer is what really caught my attention

@huggingface Natix Network is a top project of the physical AI sector.

@huggingface Well researchers have their work spelt out for them

@huggingface Huge contribution to the autonomous driving community.

@huggingface 2027 🚀🚀🚀🚀

@huggingface The combination of visual data and vehicle signals is what makes this interesting to me

@huggingface i would be curious to see what new training tasks this enables

@huggingface Telemetry alongside video makes the dataset much more useful for training.

@huggingface Open datasets like this give researchers more material to build with.

@huggingface Video shows what happened, but telemetry shows exactly how the vehicle responded. That extra layer could be huge for training.

@huggingface NATIX keeps adding more depth to the data, I’m curious what builders will create with this one.

@huggingface Third dataset already, and the data keeps getting more useful.