Detailed Program

As autonomous agents proliferate in the real world, both in software and robotic settings, they will increasingly need to band together for cooperative activities with previously unfamiliar teammates. In such “ad hoc” team settings, team strategies cannot be developed a priori. Rather, an agent must be prepared to cooperate with many types of teammates: it must collaborate without pre-coordination. This talk begins by presenting recent advances in ad hoc teamwork. The second part of the talk introduces Delta-tolling, a novel adaptive pricing scheme for multiagent traffic optimization.

Collective learning and model fusion are emerging areas that aim to pave the way forward to broaden the deployability of AI/ML in multi-agent systems. This is motivated from a limitation of existing systems which often deploy AI in narrow, offline and single-agent contexts, thus resulting in in-house/local solutions using local/in-house resources that fail to utilize the collective computation and communication capability of the entire system to support its ever-growing scale. In this talk, I will define collective learning and its key desiderata of (a) exploiting the collective resources for computation via on-demand model fusion; (b) representing local knowledge efficiently for peer-to-peer communication; and (c) developing a decentralized knowledge propagation algorithm to ensure consensus across the entire growing crowd of agents. I will also discuss the challenge of deploying collective learning in information-sensitive domains where independent agents are further constrained with the heterogeneity and privacy of their model architecture, which increases the difficulty in achieving the above desiderata. To this end, I will present our preliminary development of a set of principled solution techniques to make the first step towards achieving the above goals.

There are multi-robot motion planning (MRMP) problems involving dozens of robots, which can be speedily solved, while other are practically unsolvable. What makes an MRMP problem easy or hard? The first part of the talk will describe our quest to resolve this issue, and some progress we have made in the context of unlabeled MRMP.
The second part of the talk will review recent algorithms that we have developed for various types of MRMP problems in tight obstacle-cluttered environments: from sampling-based methods tailored to the coordination of a few complex robots, to complete and exact solutions for hundreds of simply shaped robots. The talk will conclude with surprisingly open problems for the coordinated motion of two robots.

As robots become part of the fabric of human life, devising computational models for robot interactions with other robots as well as interaction with humans is rapidly becoming an increasingly important research area. To realize this future, robots must be able to autonomously coordinate and also develop capabilities of teaming with humans. Explicit centralized control becomes increasingly impractical as robots grow in number and capabilities. On the other hand, emergent group behaviors that may result from decentralized control are particularly ill suited for humans to predict or interact with. In particular, emergent swarm performance presents challenges for human trust which may result in maladaptive human-autonomy trust calibration. This could be problematic since trust has been shown to be an important component of human-autonomy teams. In this talk, I will present our recent work in swarm coordination, human-swarm teaming, methods for assessing human trust, and optimal timing for switching between swarm behaviors

In this talk I will discuss resilience in multi-robot systems. In the first part of the talk I will attempt to clarify and disambiguate related ideas in multi-robot systems including but not limited to robustness, recovery, and resilience. In the second part of the talk (joint work with Ragesh Ramachandran and James Preiss), I will describe recent work on a method to maintain high resource availability in a networked heterogeneous multi-robot system subject to resource failures. I will describe a model in which sensing and computational resources are available on robots that are engaged in a joint task. When a resource on a particular robot becomes unavailable (e.g., a sensor ceases to function), our method automatically reconfigures the system so that the robot continues to have access to this resource by communicating with other robots. Our method can compute a communication topology, spatial formation, and formation change motion planning in a few seconds. I will show the results of our method operating in simulation and preliminary physical experiments with a small team of seven quadrotors.