Multi-Agent Coordination from Specifications
We developed techniques for coordinating large teams of heterogenous agents from temporal logic specifications. This work leverages a fragment of signal temporal logic that is chosen to be amenable to large numbers of agents. The result is a scalable framework that has been demonstrated on very large numbers of agents in simulation, as well as large teams of real robots.
More information:
Composable RL from Specifications
This work focuses on guaranteed reinforcement learning, using specifications expressed in Boolean and temporal logic. The goal is to train policies on simple, atomic tasks that can be composed at runtime with no additional training. Specifically, we focus on including both safety and reachability requirements.
More information:
Safety-Aware Task Composition for Discrete and Continuous Reinforcement Learning
Safe Control with Machine Learning in the Loop
Description coming soon.
More information:
Uncertainty Quantification for Recursive Estimation in Adaptive Safety-Critical Control
Human-Robot Tasking via Formal Specifications
This work focuses on advancing the ability of a non-expert user to task an autonomous system. While many planning formalisms exists for expressing high-level objectives, it remains difficult for most end-users to specify their goals in those formalisms. We attempt to understand how humans interpret formal languages and how we can make them more interpretable to humans. Likewise, we attempt to develop methods for eliciting expectations and preference from users that can be incorporated into planners.
More information: