> Large amounts of data from normal driving are probably good for training lane keeping, but not that useful for handling emergency situations.
The supposition I've seen is that they have models running locally to extract the most 'surprising' situations (for example, situations where the driver's inputs and the self-driving algorithm's differ substantially, or where other vehicles do not behave as expected) to send back to the mothership to train on. This would do a ton to improve their ability to deploy actual self driving in the future, and is far more than just 'normal driving' or 'lane keeping'.
The supposition I've seen is that they have models running locally to extract the most 'surprising' situations (for example, situations where the driver's inputs and the self-driving algorithm's differ substantially, or where other vehicles do not behave as expected) to send back to the mothership to train on. This would do a ton to improve their ability to deploy actual self driving in the future, and is far more than just 'normal driving' or 'lane keeping'.