The timeliness of this workshop is granted by our aspirations to solve increasingly hard problems in complex, contested, and adversarial environments. We wish to see teams of heterogeneous agents (robots and humans) operating in such unstructured environments through behaviors that are not brittle and preprogrammed, but rather adaptive, learned on-line, and resilient. The aim of this workshop is to discuss three key approaches to achieving resilient distributed intelligence and resilient behavior, in robot-robot as well as human-robot teams: composing, acting, and learning.
Juxtaposition of Three Approaches
- Abstractions and Compositionality: Heterogeneous team composition is powerful because it allows for true collaboration in changing and unpredictable environments. For example, if an ongoing task suddenly requires new capabilities, multiple heterogeneous robots can complement each other and their human teammates by performing complementary task components, or by taking advantage of differing computational and cognitive capabilities. In order to unlock this potential, we have yet to understand how to model diversity and control heterogeneous teams.
- Action and Perception: Resilient robots are capable of recognizing and rejecting unexpected, outlying, or even adversarial inputs. But as the number of robots grows, state-of-the-art strategies often become increasingly inappropriate in terms of response-time of the team, and their ability to quickly affect the team's behavior. Many tasks demand real-time response in quickly changing, adverse environments. Yet, current monolithic inference and planning architectures only work if all the data can be aggregated ahead of time. Online adaptation requires exploiting components of prior knowledge in new ways.
- Learning for Resilience: Successful performance in challenging environments must leverage data in a myriad of ways, to estimate environment geometry, topology, and semantics, to predict the behavior of teammates and other non-cooperative (and possible adversarial) entities in the world, and to infer good action strategies. Machine learning, ranging from simple system identification to techniques that use large tensor representations (e.g., deep neural network function approximation), has emerged as a key tool to automatically extract sensing models and complex perception features that provide semantic meaning suitable for high-level autonomy and learning complex control policies.
This workshop will bring together experts in these three domains. The objective of this workshop is to find connections, trigger new ideas, and foster collaborations.
- Angela Shoellig, University of Toronto, CA
- Calin Belta, Boston University, USA
- Dimitra Panagou, University of Michigan, USA
- Giovanni Beltrame, Ecole Polytechnique de Montreal, Canada
- Hadas Kress-Gazit, Cornell University, USA
- Ingmar Posner, Oxford University, UK
- Lorenzo Sabattini, UNIMORE, Italy
- Julie Shah, MIT, USA
- Mac Schwager, Stanford, USA
- Sanjit Seshia, UC Berkeley, USA
- Seth Hutchinson, Georgia Institute of Technology, USA
Topics of Interest
- Composition, control and coordination strategies for heterogeneous robot teams
- Transient response and resilient behaviors
- Machine learning and adaptation
- Analysis of resilience in robot teams
- Perception and sensor fusion in heterogeneous robot teams
- Specialization, task allocation and role assignment
University of Cambridge
Army Research Laboratory
University of Southern California
University of Pennsylvania