UC Berkeley

Despite recent advances in and many potential benefits of autonomous vehicles (AVs), therestill remain several challenges until wide-scale adoption. The largest gap is in trust, where both passengers and proximal road users need to feel comfortable using and sharing the road with AVs. To bridge this gap, both prediction and planning algorithms need to include high fidelity driver models. This will help AVs to integrate better in mixed roadways and behave in a manner that conforms to the expectations of human stakeholders, helping to improve trust and acceptance.

We investigate algorithmic frameworks that adapt to observed behaviors and balance routeefficiency with safety. After demonstrating the benefits of context-aware, data-driven, multi- modal predictions both for nominal and set-based trajectory prediction, we tackle the issue of handling prediction errors made online with a confidence-aware framework. In particular, a confidence threshold is adaptively updated online based on the consistency of behaviors with the corresponding predictions. This enables the AV to make efficient route progress when faced with a consistent target vehicle, but prioritize safety faced with an erratic target vehicle. In addition to the confidence-aware approach, we present an alternative approach that incorporates feedback policies in a stochastic model predictive control framework. Benefits in metrics of mobility, comfort, and efficiency demonstrate the advantages of adaptive feedback policies for multimodal predictions.