A single, unified AI model connecting the sensor inputs (e.g. RGB cameras, LIDAR) directly to the control output (e.g. steering, acceleration, braking). The AI learns solely from a collection of large datasets of driving experiences, whether from the real world or simulated, instead of relying on hand-coded rules. This provides the promise of greater coverage of traffic scenarios and adaptability to novel scenarios.
By learning generalized driving concepts from the data, an AV 2.0 system can adapt not only to a wider variety of nominal and emergency driving scenarios but also to a wide variety of locations. AV 2.0 systems promise shedding the need to generate and verify detailed, pre-defined HD-maps allowing rapid deployment to new or highly dynamic environments like a city under constant construction.
The crucial need of any AI system is data, especially the massive models of AV 2.0. Unlike AV 1.0 systems requiring many data formats, AV 2.0 systems rely on a single, consistent input-output data structure. The challenge is to not just capture enough data in nominal traffic conditions but also a statistically significant amount of rare, safety-critical edge cases for complete model training, validation, and verification. Conventional data logging is very expensive and does not guarantee coverage. Photorealistic and human-realistic synthetic data provides a cost effective solution for filling gaps within datasets.