Harnessing Human Intelligence for Adaptive Human-Robot Collaboration (white paper tracking # 20-000000167)
Abstract
Humans are capable of seamlessly collaborating with one another in teams to solve complex challenges, including advanced manufacturing and construction. In many settings, autonomous robots are also expected to collaborate with humans to achieve shared goals. However, even the most advanced robots today fall short of these expectations. The goal of this proposal is to learn sharedrepresentations from successful human-human collaboration to develop new algorithms for seamless, efficient, and adaptable human-robot and multi-robot collaboration.Objective 1: Learning Shared Representations. Collaborators need to share a common representation of their shared goals, as well as a representation of one anothers internal states. For example, when assembling a table, agents should employ an object-oriented representation of the relevant physical entities (i.e., tabletop, legs) and their spatial relationships, and infer which objectsare visible and accessible from their partners current vantage point. Our project will create and publicly release benchmark datasets for human performance on such assembly tasks, thus exposing the low-dimensional representations that humans employ to perform these tasks. Guided by insights from these empirical studies, our project will develop new algorithms that learn to encode shared low-dimensional task representations that mirror those used by humans.Objective 2: Learning to Coordinate Actions. Collaborators need to be able to predict each others actions to successfully coordinate. Next, we will develop and empirically evaluate algorithms that learn to predict the most likely policy adopted by a partner conditioned on their previous actions, supporting an appropriate and complementary policy in response. For example, duringtable assembly one agent must stabilize the tabletop while another agent attaches the next leg. As such, observing that ones collaborator has reached for the next leg supports the inference that it would be appropriate to stabilize the tabletop. The ability to predict a partners future actions introducesa set of new opportunities for developing adaptive robots that can even positively influence a partner toward more beneficial outcomes, for example, by taking actions that guide the human to take on complementary and mutually beneficial roles and subtasks.Objective 3: Continual Learning and Adaptation. Collaborators need to learn how to work together more effectively over time, as well as adapt to changing task conditions in a coordinated manner. For example, after assembling multiple tables together, both agents should update their expectations about the role that each agent will adopt (i.e., stabilizing tabletop v.s. attaching leg), and be able to generalize these expectations to novel assembly tasks (e.g., assembling other furniture) and new collaborators. Building on Objectives 1 & 2, we will develop and empirically validate algorithms that not only learn to collaborate and adapt to their partners more efficiently on one task after repeated interactions, but also generalize quickly to novel collaboration contexts.Impact. Our project will develop new algorithms in concert with controlled human experiments that characterize collaborative behavior in a detailed, quantitative manner. This tight integration between model development and empirical evaluation has strong potential to both advance the state of the art in human-AI and human-robot interaction and to produce quantitatively precise psychological theories of social coordination in complex and changing task environments.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 09, 2021
- Source ID
- N000142112298
Entities
People
- Dorsa Sadigh
Organizations
- Office of Naval Research
- Stanford University
- United States Navy