The primary challenge in autonomous driving is identifying and collecting data on rare, anomalous scenarios that models are not trained to handle.
Cariad, a Volkswagen subsidiary, is building a unified software and tech stack to accelerate innovation for all Volkswagen group brands.
The Big Loop system solves the high cost of traditional data collection by intelligently aggregating only useful information using dedicated hardware.
The long-tail problem refers to rare but critical events, like noisy sensor data or unknown objects, that can be identified using methods like uncertainty estimation.
The INSTINCT software uses deep neural networks to analyze sensor data in real-time and calculate an uncertainty score to flag challenging scenarios for collection.
The Big Loop enables a continuous cycle of driving, uploading valuable data, labeling, retraining models, and deploying them back to vehicles via over-the-air updates.
The Big Loop technology is being deployed in a retrofitted "pioneering fleet" to scale data collection before its full rollout in millions of future vehicles.
The discussion covers ethical dilemmas, local vs cloud model execution, setting dynamic uncertainty thresholds, and the goal of creating globally applicable models.
52:10Nico Schmidt
48:01Jan Zawadzki
26:55Jan Zawadzki
1:00:28Tillman Radmer & Fabian Hüger & Nico Schmidt
31:23Sayed Bouzouraa
38:12Daniel Graff & Andreas Wittmann
42:18Francis Powlesland & Elena Kotljarova
35:16Ankit Patel
Jobs that call for the skills explored in this talk.
DTS Systeme GmbH
Transportation and Infrastructure Systems
Ingolstadt, Germany
Universität Bern