* Internal feature space information content increased tenfold during training (to ~700 bits), this makes the model dramatically more accurate
* Internal feature space information content increased tenfold during training (to ~700 bits), which makes the model dramatically more accurate
* Less reliance on previous frames makes model more reactive and snappy
* Less reliance on previous frames makes model more reactive and snappy
* Trained in new reprojective simulator
* Trained in new reprojective simulator
* Model trained in openpilot was trained in 36hrs from scratch, compared to around 1 week of previous releases
* Trained in 36hrs from scratch, compared to one week for previous releases
* Model training now simulates lateral and longitudinal behavior, this allows openpilot to slow down for turns, stop at traffic lights, etc,... in experimental mode
* Training now simulates both lateral and longitudinal behavior, which allows openpilot to slow down for turns, stop at traffic lights, and more in experimental mode
* New driver monitoring model
* New driver monitoring model
* New end-to-end distracted trigger
* New end-to-end distracted trigger
* Experimental driving mode
* Experimental driving mode
* End-to-end longitudinal control
* End-to-end longitudinal control
* Stops for red lights and stop signs
* Stops for traffic lights and stop signs
* Slows down for turns
* openpilot defaults to chill mode, enable experimental in settings
* openpilot defaults to chill mode, enable experimental in settings