Davis S. Catherman¹, Ashay Aswale¹, Steve Marotta², Spencer K. Lynn², David Koelle², and Carlo Pinciroli¹

Poster Session at the Lunar Surface Innovation Consortium (LSIC) (13-15 November 2024)

Introduction

Multi-robot collaboration and interaction are critical autonomous technologies to enable future lunar missions [1]. In 2025, NASA’s Cooperative Autonomous Distributed Robotic Exploration (CADRE) mission will be an early technology demonstration of three rovers working together on the lunar surface to collaboratively map an environment [2]. Looking forward, as the number of robots scales, so will the need for increased autonomy in individual control and inter-robot interactions. To best coordinate actions, robots should have access to global knowledge as frequently as possible. However, as robots spread in the environment, the quality of the available knowledge on each robot is limited by the frequency of robot interactions and access to centralized base stations. In the absence of a global satellite network, creating an effective network infrastructure for extraterrestrial missions remains a considerable challenge.

To overcome these issues, we study how to endow robot swarms with a theory of mind [3] that enables a swarm to predict, reason, and plan even when communication is infrequent. The basic intuition is that the robots possess predictors for relevant mission-critical variables. These predictors compensate for periods in which the robot is isolated. The ground truth for the predicted variables is updated in a best-effort fashion when connected to the base station or meeting other robots.

The benefits of realizing such theory of mind (ToM) are twofold: (i) the robots can reason and plan about collaborative tasks in the absence of communication; (ii) the predictions act as a system model that enables autonomous fault detection and diagnosis.

Our approach is a novel way to tackle the data Consistency, data Availability, and Partition tolerance from the CAP theorem [4], which states that only two of the three elements can be guaranteed at once when sharing data. In contrast to existing work, our ToM is based on offering partition tolerance and consistency through prediction capabilities, and availability through the output of the predictors.

In our work, we first focus on characterizing the peer-to-peer interactions, which we refer to as occasional interactions. Subsequently, we use these occasional interactions to fulfill a knowledge transfer design capable of maintaining knowledge accuracy over long periods of isolation. Through the development of this work, we aim to enable future lunar missions that require multiple robot collaborative autonomy over large areas.

Occasional Interaction Characteristics: We observe the occurrence of occasional interactions when, on average, robots are separated by distances significantly larger than their communication range, and their velocity leads to long traversal times to reach another robot. From this analysis, we identify large times (τ) that are characteristic of occasional interactions due to the long traversals and slow rate of mixing. Additionally, the duration of each interaction is sufficiently long, allowing full knowledge transfer upon meeting. Our modeling, which pessimistically denies explicit rendezvous of nearby robots, demonstrates the ability for the robots to still propagate data over multiple time periods defined by τ. Future works may also find such modeling beneficial for selection of hardware as it relates to system design.

Theory of Mind (ToM) Knowledge Transfer: We crafted a ToM structure that proactively provides consistency via Conflict-free Replicated Data Types (CRDTs) [5]. CRDTs are shared data structures in which conflicts are automatically solved upon merging of data. This enables satisfying eventual consistency across the ordered history of knowledge, effectively providing global knowledge for up-to-date robots.

We also find novel use from the data that is eventually shared across all robots. The same data that allows orderable storage and conflict resolution provides estimation and prediction capabilities; we effectively get two benefits without increasing storage costs. Our early evaluation demonstrates improvement to understanding the state of other robots such as estimation from health parameters, yielding ToM capabilities. The dynamics also show improvement in task prioritization consensus amongst multiple robots.

For More Information

To learn more, contact Spencer Lynn or David Koelle.

(Please include your name, address, organization, and reference. Requests without this information will not be honored.)