Modeling human driving from demonstrations
Abstract/Contents
- Abstract
- For autonomous agents to coexist and cooperate with humans, it is important for them to anticipate human behavior. Models of human behavior can be used by autonomous agents to plan in response to human actions and proactively coordinate with humans. Such models can also be used to build simulators for testing autonomous agents by replicating the environment of operation. Modeling human behavior is challenging because of multiple reasons such as stochasticity, multi-modality, unobservable intents, high dimensional state action spaces, and nonlinear dynamics. In the literature, both ontological and phenomenological approaches have been used to model human behavior. While ontological approaches use rules, phenomenological approaches are data-driven. This thesis presents techniques to model human behavior from demonstrations. The techniques proposed in this thesis are evaluated on their ability to model human driving behavior which is important in autonomous driving for both planning and safety validation. First, this thesis adapts the technique of Generative Adversarial Imitation Learning to the problem of driver modeling. It extends the GAIL formulation to work in the multi-agent setting where observations gathered from multiple agents are used to inform the training process of a learning agent. The proposed method is shown to better imitate demonstrated driving as opposed to single agent learning method. Since driving has associated rules, the second part of this thesis introduces a method to provide domain knowledge to the imitation learning agent through reward augmentation. The proposed method, which relies on reward augmentation, is shown to provide better emergent driving performance and overall traffic flow in the recreated traffic simulations. Many of the applications of autonomous agents are safety-critical including that of autonomous driving. This makes it important for models to be interpretable. This thesis proposes a hybrid rule-based and data-driven method that relies on the technique of particle filtering to learn parameters of underlying rule-based models from human driving demonstrations. The proposed method is demonstrated on the problem of highway merging and shown to generate realistic driving behavior as assessed by a driving Turing test.
Description
| Type of resource | text |
|---|---|
| Form | electronic resource; remote; computer; online resource |
| Extent | 1 online resource. |
| Place | California |
| Place | [Stanford, California] |
| Publisher | [Stanford University] |
| Copyright date | 2021; ©2021 |
| Publication date | 2021; 2021 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Author | Bhattacharyya, Raunak Pushpak |
|---|---|
| Degree supervisor | Kochenderfer, Mykel J, 1980- |
| Thesis advisor | Kochenderfer, Mykel J, 1980- |
| Thesis advisor | Sadigh, Dorsa |
| Thesis advisor | Schwager, Mac |
| Degree committee member | Sadigh, Dorsa |
| Degree committee member | Schwager, Mac |
| Associated with | Stanford University, Department of Aeronautics and Astronautics |
Subjects
| Genre | Theses |
|---|---|
| Genre | Text |
Bibliographic information
| Statement of responsibility | Raunak Bhattacharyya. |
|---|---|
| Note | Submitted to the Department of Aeronautics and Astronautics. |
| Thesis | Thesis Ph.D. Stanford University 2021. |
| Location | https://purl.stanford.edu/ws309yz7055 |
Access conditions
- Copyright
- © 2021 by Raunak Pushpak Bhattacharyya
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...