User Tools

Site Tools


8a.005.UVA - Improved Decision Making for Autonomous Systems

Project - Summary

Ships, or vessels, often sail in and out of cluttered environments over the course of their trajectories. Safe navigation in such cluttered scenarios requires an accurate estimation of the intent of neighboring vessels and their effect on the self and vice-versa well into the future. In manned vessels, this is achieved by constant communication between people on board, nautical experience, and audio and visual signals. In this project we propose a deep neural network based architecture to predict intent of neighboring vessels into the future for an unmanned vessel solely based on positional data.

Project - Team

Team Member Role Email Phone Number Academic Site/IAB
Cody Fleming PI (434) 924-7460 University of Virginia
Steve Adams Co-PI (757) 870-4954 University of Virginia
Jasmine Sekhon Student Not available Not available University of Virginia
Mike Raker Project Mentor Not available Leidos
Funded by: Leidos

Project - Novelty of Approach

The usability of data-driven decision-making solutions in safety-critical applications is limited by the absence of human-like inference and reasoning about decisions. Moreover, most models overlook the possibility of more than one feasible decision and end up optimizing on average behavior. This project is expected to improve decision-making models by incorporating human expert knowledge and interpret ability in the model. The project will focus on maritime domain that has different conditions from land or air based research projects focusing on autonomous agents.

Project - Deliverables

1 Preliminary model design
2 Literature review
3 Data collection
4 Prototype model

Project - Benefits to IAB

In this work, we propose a trajectory prediction framework for socially interacting agents, that we evaluate on maritime datasets. Broadly, our framework is general enough to be applicable to any other kinds of agents, for example, urban road traffic prediction, intent prediction for pedestrians in crowded environments, or any other multi-agent setting. This work can also be extended to predicting multiple socially plausible trajectories per agent in the scene to account for the multimodal nature of navigation. While the performance of this model is better than most state-of-the-art trajectory prediction methods, more performance benefits can be achieved by including additional input features, such as scene information, vessel type, etc. This work can also be extended to predicting trajectories for heterogeneous agents with different trajectory dynamics. The spatial attention mechanism introduced in this work can be used to infer more domain-specific knowledge, usch as the influence of different kinds of agents on each other (for example, the effect of a skateboarder on a cyclist's trajectory) and use these to either model predictions or inform model predictions.

At a more fundamental level, our approach is a general framework that can be applied to any sequence-to-sequence modeling application where cross-LSTM knowledge can help improve performance. This can include human action recognition, modeling human-object interaction, video classification. An important advantage of the approach predictions to infer domain knowledge from the observation dataset and hence yield improved predictions without making significant assumptions about the application domain or the dataset.

Project - Documents

FilenameFilesizeLast modified
8a.005.uva_final_report.pdf861.5 KiB2020/06/30 16:19
8a.005.uva_mid-year_report.docx239.8 KiB2019/12/20 11:15
8a.005.uva_year_8_pitch.pptx123.9 KiB2019/09/18 17:27
8a.005.uva_year_8_project_proposal_cody_fleming.pdf686.6 KiB2019/09/18 17:25

Life Form Feedback

Year 8 Project Poster Session Feedback (Fall 2019)
Real name Great Progress On Course Needs Change Off Course Abstain
Kimmo Valtonen (kimmo.valtonen)     
Sumit Shah (sumit.shah)     
projects/year8/8a.005.uva.txt · Last modified: 2021/06/02 17:23 by sally.johnson