Learning, Leveraging, and Influencing Representations for Interactive Autonomy
Abstract
Approved for Public ReleaseTitle: Learning, Leveraging, and Influencing Representations for Interactive AutonomyPI: Dorsa Sadigh, St,anford University, Budget: $ 510,000Research Problem: Machine learning has made significant advances in developing autonomous and in,telligent systems; however, most algorithms are still designed for agents acting in isolation. In practice, interaction with humans,and other learning agents is inevitable in advancing the science of autonomy. Such interactions present a significant set of challen,ges: the other agents will update their behavior in response to the autonomous ego agent, continually changing the learning environm,ent of the agent. The core of our research problem is to answer the following questions: How can we predict future behaviors, given,that other agents continuously change and adapt over time? How do other agents respond and adapt to interventions, and can autonomou,s and intelligent agents steer the system toward more desirable behaviors?Research Plan: Our key insight is that complex interaction,s can be succinctly predicted once we use the right representations. This proposal will take representation learning ideas that have, been at the core of the deep learning revolution, and extend them in the context of non-stationary multi-agent systems to automatic,ally learn and leverage representations for adaptive interaction.Technical Approach: We will address key gaps in the state-of-the-ar,t to improve the ability to interact with and influence complex multi-agent systems. Specifically, our approach is as follows:1) Lea,rning and leveraging latent representations for coordination: We plan to develop unsupervised learning techniques that directly lear,n representations in non-stationary environments. These representations can then be used for the goal of coordination and collaborat,ion with other agents.2) Formalizing structures in other agent policies: Formalizing priors such as low-dimensionality, low-rankedne,ss, sparsity, or graph networks can enable us to develop scalable representation learning techniques that capture structures in the,strategy of other agents or teams, which can help with better anticipating their actions.3) Influencing interactions: We will go bey,ond simple one-step coordination, and study how our actions can influence other agents over multiple interactions. Specifically, we,will develop algorithms that influence other agents in cooperative or competitive multi-agent settings.Expected Outcome: The expecte,d outcome of this project is new algorithms, tools, and analysis of learning-based techniques for partner modeling in multi-agent sy,stems. This includes algorithms that both interact with other intelligent autonomous agents or humans. We expect this project to lea,d to research publications, open-source software, and benchmarks that can enable standardizing multi-agent learning problems in both, cooperative and competitive scenarios.Impact on DoD Capabilities: If successful, this project will provide new theoretical and algo,rithmic tools for partner modeling, coordination, and influencing in multi-agent interactions. This can potentially impact many mili,tary capabilities and naval applications such as coordination between underwater vehicles for search and rescue missions, collaborat,ion between humans and autonomousagents on naval ships for tasks such as repair, or assistive teleoperation of autonomous robots for, munition handling. The outcomes of this project will be a fundamental step toward developing interactive and adaptive autonomous sy,stems in military applications.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 16, 2022
- Source ID
- N000142212293
Entities
People
- Dorsa Sadigh
Organizations
- Office of Naval Research
- Stanford University
- United States Navy