Skip to main content
eScholarship
Open Access Publications from the University of California

Interactive Prediction and Planning for Autonomous Driving: from Algorithms to Fundamental Aspects

  • Author(s): Zhan, Wei
  • Advisor(s): Tomizuka, Masayoshi
  • et al.
Abstract

Inevitably, autonomous vehicles need to interact with other road participants in a variety of highly complex or critical driving scenarios. It is still an extremely challenging task even for the forefront companies or institutes to enable autonomous vehicles to interactively predict the behavior of others, and plan safe and high-quality motions accordingly. The major obstacles are not just originated from prediction and planning algorithms with insufficient performances. Several fundamental problems in the fields of interactive prediction and planning still remain open, such as formulation, representation and evaluation of interactive prediction methods, motion dataset with densely interactive driving behavior, as well as interface of interactive prediction and planning algorithms.

The aforementioned fundamental aspects of interactive prediction and planning are addressed in this dissertation along with various kinds of algorithms. First, generic environmental representation for various scenarios with topological decomposition is constructed, and a corresponding planning algorithm is designed by combining graph search and optimization. Hard constraints in optimization-based planners are also incorporated into the training loss of imitation learning so that the policy net can generate safe and feasible motions in highly constrained scenarios. Unified problem formulation and motion representation are designed for different paradigms of interactive predictors such as planning-based prediction (inverse reinforcement learning), as well as probabilistic graphical models (hidden Markov model) and deep neural networks (mixture density network), which are utilized for the prediction/planning interface design and prediction benchmark. A framework combing decision network and graph-search/optimization/sample-based planner is proposed to achieve a driving strategy which is defensive to potential violations of others, but not overly conservatively to threats of low probabilities. Such driving strategy is achieved via experiments based on the aforementioned interactive prediction and planning algorithms with proper interface designed. These predictors are also evaluated from closed loop perspective considering planning fatality when using the prediction results instead of pure data approximation metrics. Finally, INTERACTION (INTERnational, Adversarial and Cooperative moTION) dataset with highly interactive driving scenarios and behavior from international locations is constructed with interaction density metric defined to compare different datasets. The dataset has been utilized for various behavior-related research areas such as prediction, planning, imitation learning and behavior modeling, and is inspiring new research fields such as representation learning, interaction extraction and scenario generation.

Main Content
Current View