Lecture, Presentation
Drago Anguelov (Waymo) - MIT Self-Driving Cars
- Full removal of human operators is anticipated for self-driving vehicles, though large-scale deployment is projected to take decades rather than the next couple of years due to the need for extensive logistics, algorithm tuning, and safety verification.
- Taming the "long tail" of rare scenarios is identified as a critical requirement for safety, with this scope expected to expand as services enter new environments featuring unique challenges like complex intersections and varying local driving customs.
- System planning and prediction will need to anticipate agent behaviors over durations ranging from one second to potentially ten seconds or more, while also accounting for non-standard cues such as waving or distracted individuals.
- Future architectures will likely evolve from a hybrid approach toward greater reliance on machine learning and reduced dependence on expert-designed rules, with the system designed to drive conservatively when prediction confidence is low.
- Capabilities to quantify uncertainty, detect inconsistency in beliefs, and leverage environmental constraints like temporal sequences will be essential for enabling actions and improving model robustness over time.
- Active learning and simulation environments will be utilized to select and label rare, uncertain cases, requiring realistic modeling of driver and pedestrian behavior, including adversarial agents and the mapping of simulator scenes to real-world conditions.
- Current reaction models such as "break and swerve" are expected to be insufficient for complex interactive scenarios like merges and lane changes, necessitating a mix of models tuned to various aspects of human behavior distributions.
- Progress relies on the development of smart agents that facilitate better decision-making and prediction, alongside reasoning methods like consistency checks between ego motion and depth estimation to support long-term model improvements.