A Prediction and Planning Framework for Scalable Autonomous Driving in Urban Areas

Abstract

In the past few years, automotive and technology companies have made major progress towards the real-world deployment of autonomous driving technologies. A few companies have launched fully autonomous driving technologies for taxi rides in small geographic areas. With the initial milestone made, the challenge of developing effective and scalable autonomous driving technologies that can operate in a wide variety of complex urban environments has never been more important.

I propose a prediction and planning framework for self-driving in urban areas to address the challenge. With particular attention to the scalability of the approach, the framework considers unique contexts to each environment and generates effective trajectory plans for different variations of urban driving scenarios in a computationally efficient way. The framework consists of two main tasks: prediction of the environments and planning trajectories of the autonomous vehicle. I mainly leverage learning-based techniques, which have experienced significant progress in recent years, for both prediction and planning tasks.

The prediction task is critical in the framework as accurate predictions of the environment states and their uncertainties are vital to safe and optimal decision-making. To this end, I introduce two powerful conditional generative models, namely HCNAF and CVAE-H, based on normalizing-flow and variational autoencoder, respectively. I show that the two algorithms effectively leverage social and spatial sensor information such as past trajectories of the road-agents and lidar scans of the environment for forecasting the motions of the road agents in a diverse set of environments. I compare my prediction models against state-of-the-art methods using an urban driving dataset and show both methods achieve improved prediction accuracy.

I design the planner to generate near-optimal action sequences autonomously and to consider the uncertainties captured in the prediction outputs. The proposed planner is a model-based random shooting planner with a Gaussian mixture as the backbone distribution. The Gaussian mixture is parameterized using a deep neural network and trained using cross-entropy loss and rewards of the sampled trajectories. Experiments confirm that the proposed planner generates contextual trajectories under various environments in real-time, and the performance compares favorably against several baseline planners, including a dynamic programming planner.

Lastly, I compare the computational efficiency of two different uncertainty representations of the environment. The two representations are (1) trajectory samples of road agents and (2) probabilistic occupancy map, which encodes occupancy probabilities of the road-agents on a continuous 2D heat map. I examine their performances and show that the probabilistic occupancy map representation offers faster and more scalable inference without an excessive sampling of the future states of the road-agents.