An Integrated System for Multi-Rover Scientific Exploration

Tara A. Estlin, Alexander G. Gray, Tobias Mann, Gregg R. Rabideau, Rebecca Castaño, Steve A. Chien, Eric MjolsnessAAAI 1999

Abstract

This paper presents the Multi-Rover Integrated Science Understanding System (MISUS), an architecture designed to coordinate multiple planetary rovers through integrated machine learning and autonomous planning. By performing on-site spectral data analysis and autonomous goal generation, MISUS enables real-time science exploration with minimal human intervention, addressing the significant communication latencies inherent in deep-space missions.

Introduction

Planetary exploration has historically relied on "human-in-the-loop" operations, where data is transmitted to Earth, analyzed by scientists, and then used to script specific commands for the next mission cycle. With signal delays reaching 20 minutes for Mars and limited communication windows, this process is inefficient for multi-rover fleets. MISUS addresses this by closing the loop between science understanding (machine learning) and mission execution (planning and scheduling).

System Architecture

The MISUS framework integrates distributed data analysis with autonomous planning to manage a fleet of rovers. The architecture is divided into central and local components to balance global coordination with local resource management.

The MISUS architecture integrates central data analysis and planning with individual rover execution and clustering.

As shown in Figure 1, the system consists of three primary pillars:

  1. Distributed Data Analysis: Rovers perform local computations on spectral data and share model parameters with a central node to build a global geological map.
  2. Autonomous Goal Selection: A heuristic system identifies scientific targets based on the global model, prioritizing both representative samples and rare anomalies.
  3. Distributed Planning: The ASPEN (Automated Scheduling and Planning Environment) framework allocates these goals across the fleet.

Science Understanding and Data Analysis

The system utilizes an unsupervised clustering algorithm based on Expectation-Maximization (EM). Instead of transmitting high-bandwidth raw spectral images, rovers transmit small sets of model parameters. This approach is highly bandwidth-efficient, a critical requirement for space missions.

Clustering of rock spectra features into distinct geological categories based on frequency bins.

Figure 2 demonstrates how the system categorizes rock types in a feature space derived from 14-dimensional spectral readings. By identifying these clusters, the system can autonomously determine which areas of the terrain warrant further investigation without waiting for human input.

Mission Planning and Execution

Mission planning is handled by the ASPEN framework, which treats goal allocation as a Multi-Traveling Salesman Problem (MTSP). A central planner divides goals among rovers to minimize total travel time. Individual rovers then use "iterative repair" algorithms to generate detailed action sequences that respect local constraints such as battery life and instrument availability. This distributed approach ensures resilience; if one rover fails, its goals can be re-allocated to the remaining fleet.

Results and Conclusion

Simulation results using Mars-like "rockscapes" demonstrated that MISUS could effectively build accurate geological models through minimal data exchange. The system successfully reduced total traversal distance and showed the ability to autonomously re-task rovers upon identifying new rock types. While the 1999 prototype used simplified physical models, it laid the foundational logic for modern autonomous behaviors seen in missions like the Mars 2020 Perseverance rover.

Summary updated

BibTeX
@inproceedings{estlin1999integrated,
  author = {Tara A. Estlin and Alexander G. Gray and Tobias Mann and Gregg R. Rabideau and Rebecca Castaño and Steve A. Chien and Eric Mjolsness},
  title = {{An Integrated System for Multi-Rover Scientific Exploration}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1999},
}
Figure