The self-driving data flywheel runs on multiplication

At an event unveiling an autonomous-driving development and operations system, the key idea singled out is a framing: the speed of the data flywheel is determined not by the sum but by the product of scale, quality, efficiency and operations, so if even one of them is close to zero, the whole thing stops. More than 40 collection vehicles each gather data for 80 hours a week, but nearly half of it is straight-line, constant-speed driving with little training value, which leads to the diagnosis that hard mining, automatically selecting only the samples the model finds difficult, is needed. The logging system uses three triggers (manual requests, detection of sudden acceleration or braking, and disengagement of autonomous driving) to clip only the 15 seconds before and after an event, storing sensor data and even the model's intermediate outputs together, and it handles annotation automatically as well. Testing is done in a closed loop reconstructed from real driving data with 3D Gaussian Splatting, and a model on the order of 100 million parameters is quantized to fit within a latency of a few milliseconds. However, keep in mind that this is company-reported, and there is not a single comparable performance figure such as an intervention rate.

This is a note: my own write-up after reading several sources. It has no source link; events checked against a source are kept as records.