Capability Augmentation for Heterogeneous Dynamic Teaming with Temporal Logic Tasks

TL;DR

Capability-Augmenting Tasks (CATs), which encode how agents can augment their capabilities based on interactions with other teammates are presented, which integrates CAT into the semantics of Metric Temporal Logic (MTL), which defines individual spatio-temporal tasks for all agents.

Abstract

This paper considers how heterogeneous multi-agent teams can leverage their different capabilities to mutually improve individual agent performance. We present Capability-Augmenting Tasks (CATs), which encode how agents can augment their capabilities based on interactions with other teammates. Our framework integrates CAT into the semantics of Metric Temporal Logic (MTL), which defines individual spatio-temporal tasks for all agents. A centralized Mixed-Integer Program (MIP) is used to synthesize trajectories for all agents. We compare the expressivity of our approach to a baseline of Capability Temporal Logic Plus (CaTL+). Case studies demonstrate that our approach allows for simpler specifications and improves individual performance when agents leverage the capabilities of their teammates.

Figures (4)

• Figure 1: Example: An environment with two ground agents and an aerial agent. The aerial agent must take a picture in "Scenic" nodes and upload it in "Upload" nodes. The ground agent must reach "Goal" and avoid "Water". In this case, the ground robots can also upload pictures, and the aerial robot can carry one ground robot over the water at a time.
• Figure 2: Case Study 1: The agents must pick up supplies in the "Supply" node and deliver them to "Village 1" and "Village 2". Ground robots must avoid "Water" nodes and cannot visit the "Bridge" node until it is inspected by an aerial agent. Only two ground agents can visit "Bridge" at a time
• Figure 3: Example solution for the three agent case of case study 2. We show the trajectory for the ground agents' using the rectangles and show the aerial agent's trajectory with ovals. In each rectangle and oval, we specify the exact time steps that the agent occupied the node. We see the aerial agent carry one ground agent over the water at time step 1, then the other at time step 2. The aerial agent takes pictures at time steps 5 -7 and uploads an image using the ground agent at time step 10. The ground robots reach their goal nodes at time steps 2 and 3.
• Figure 4: Case Study 2 Results. Each environment size was tested 20 times with three agents in random initial nodes and randomized labels. We randomized labels so that roughly $60\%$ of the nodes were "Water", $1\%$ of the nodes were "Upload", $1\%$ of the nodes "Scenic", and $1\%$ of the nodes were "Goal". Environments are required to contain at least one of each label. The two formulations were tested in the same random environments. Here, larger agent performance indicates that an agent is better able to satisfy a task. In this case, negative agent performance can only be achieved when the majority of agents in a trial do not satisfy their individual specifications.

Theorems & Definitions (3)

• Definition 1: Syntax of CATs
• Definition 2: Qualitative Semantics of CAT
• Remark 1