Lectures, Conference Presentation
Karl Iagnemma & Oscar Beijbom (Aptiv Autonomous Mobility) - MIT Self-Driving Cars
- Aptiv anticipates autonomous driving software will constitute only one component of future transportation systems, aiming to industrialize technology to operate reliably for hundreds of thousands of miles and plans to demonstrate vehicles running 20 hours daily, seven days a week in Las Vegas.
- The industry has accumulated approximately 12 to 14 million miles of autonomous driving under hundreds of code versions, with expectations to supplement physical driving with extensive simulation and regression testing on supersized datasets to build statistical safety arguments without requiring hundreds of millions of miles on a single code build.
- Regulatory uncertainty remains a significant known unknown, characterized by a lack of consensus on safety targets across different jurisdictions like Las Vegas, Singapore, or San Francisco, and while current federal guidelines are evolving, the automotive sector has historically relied on self-certification rather than external oversight.
- Validation challenges include the difficulty of proving safety for rare accidents—requiring 275 million crash-free miles per RAND statistics to exceed human safety with 95% confidence—and the complexity of updating code continuously, which necessitates expensive and lengthy re-validation processes.
- Technical approaches are shifting away from single "black box" architectures toward modular systems where neural networks are caged within safety layers to allow rigorous validation of sub-modules and satisfy functional safety and SOTIF requirements.
- Oscar's presentation details the PointPillars method, which achieves inference speeds of 1.3 milliseconds and runs at over 60 Hz, outperforming previous metrics on the KITTI dataset, alongside the upcoming release of the NuScenes benchmark dataset containing 1,000 20-second scenes collected over a 360-degree sensor suite.
- The NuScenes dataset, set for release in approximately one month, is designed as an academic effort to address the limitations of older datasets like KITTI by providing full 3D views and radar data, though it is explicitly not claimed to be sufficient for an industrial safety case.
- Data sufficiency and bias are critical constraints, with the expectation that training data must cover all operational conditions—such as nighttime driving or specific weather events like Singapore's torrential rain and construction zones—to avoid deployment failures in unseen scenarios.
- Fleet connectivity requirements vary by use case, ranging from relaxed needs for simple position sharing to stringent high-bandwidth, low-latency demands for teleoperation, though the core system is designed to function without relying on the future 5G rollout.
- Environmental adaptation is managed through a flexible "rule book" structure that allows for the reprioritization of rules and weights to accommodate different driving cultures and regulations, while route selection and service rollouts may be phased, such as excluding nighttime operations until comfort levels are achieved.
- Interpretability is expected to become a regulatory requirement for explaining specific algorithmic decisions in the event of crashes, particularly for lay juries, as current methods lack the ability to fully explain why specific neural network outputs occur.
- LiDAR data offers distinct augmentation capabilities compared to visual data, allowing for the insertion of photorealistic objects and snow conditions into training sets more easily, though deployment is strictly limited to environments where the training data supports the specific conditions.
- The industry is moving toward treating engineering more like the natural sciences, creating complex functional approximators and investigating their properties rather than defining closed-form functions, with future self-certification likely evolving to include third-party regulatory oversight similar to nuclear or aerospace sectors.