Loading...
Publication
Multi-agent exploration of unknown terrain: Combining neural time series forecasting with multi-agent coordination for planetary surface exploration
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 39.83 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
This work addresses the challenge of autonomous robotic exploration in uncertain planetary
environments, such as the Martian surface, where long communication delays with
Earth make local decision-making essential. The main objective was to develop a simulation
framework aimed at enhancing the efficiency and autonomy of multi-agent robotic systems,
with a particular emphasis on scalable cooperation. The study investigates how the integration
of Multi-Agent Reinforcement Learning (MARL) algorithms and dynamic environmental
simulations can improve exploration in unknown terrains.
The methodology involved the development of a detailed 2D simulation environment of
Jezero Crater, Mars, incorporating realistic weather conditions modeled from the MEDA
dataset. Climate forecasting relied on N-BEATS and LSTM models, with N-BEATS demonstrating
superior performance in predicting environmental variables. Coordination between
Explorer agents (tasked with mapping and sample identification) and Transporter agents (responsible
for collection and delivery) was managed through an Auction-Based Coordination
and Autonomy Layer powered by MARL.
The results demonstrate the effectiveness of the multi-agent approach, achieving a successful
coordination rate of 79.4% across multiple simulation runs under varying atmospheric
conditions (with successful coordination defined as task assignments completed within the
mission time budget via the auction-based mechanism). Furthermore, emergent behavioral
specialization was observed among Transporter agents, with significant diversity in the confidence_
bias parameter (coefficient of variation, CV = 0.256), enhancing overall system
robustness. Additionally, the system exhibited linear scalability and a 192.9% improvement
in territorial coverage in multi-agent configurations compared to single-agent setups.
In conclusion, the integration of MARL with dynamic environmental simulations provides a
robust and adaptable solution to the complex challenges of robotic exploration and navigation
in extreme environments, paving the way for future autonomous space missions.
Description
Keywords
Multi-Agent Reinforcement Learning Autonomous Robotic Exploration Surface of Mars Multi-Agent Coordination Time Series Forecasting Dynamic Environmental Simulation