Marsplanbench and MoonPlanBench Datasets Advance Rover Navigation Technology

Planetary rovers are poised for a resurgence with planned missions to the Moon and beyond, yet valuable data from previous explorations remains largely untapped for advancing autonomous navigation technologies. Marvin Chancán, Avijit Banerjee, and George Nikolakopoulos, all from Luleå University of Technology, Sweden, address this gap by introducing the first large-scale, standardized datasets for rover path planning research. Their work establishes MarsPlanBench and MoonPlanBench, built from high-resolution terrain imagery, and provides a unified framework for evaluating both traditional and cutting-edge, learning-based path planning algorithms. This research delivers crucial new insights into the performance of these algorithms on realistic planetary surfaces, demonstrating that established methods can achieve remarkably high success rates even in challenging terrains, while highlighting the ongoing difficulties faced by machine learning approaches when applied to these complex environments. By making their datasets and code openly available, the team provides a vital resource to accelerate progress in the field of planetary robotics and autonomous exploration.

Planetary Rover Path Planning Benchmarks and Datasets

This research introduces PathBench, a comprehensive benchmarking platform for planetary rover path planning algorithms, alongside new datasets generated from real lunar and Martian terrain data. Planning efficient and safe paths for rovers presents challenges due to complex terrains and limited visibility, and existing benchmarks lack realism and scope. The team created datasets based on high-resolution terrain data, evaluating both classical algorithms like A* and RRT-Connect, and learned methods such as Waypoint Planning Networks and Value Iteration Networks. The evaluation considers path length, computation time, success rate, and smoothness, utilizing a photorealistic simulator to ensure realistic testing conditions.

Results demonstrate that Dijkstra’s Algorithm generally performs well in path optimality, while RRT-Connect balances speed and solution quality. Learned methods show promise but require substantial training data and struggle with generalization to unseen terrains, with performance varying significantly based on terrain type and algorithm parameters. This work highlights the importance of realistic simulation for robust algorithm evaluation. The study provides a standardized benchmark for comparing algorithms, offers realistic datasets for training and evaluation, and identifies the strengths and weaknesses of different approaches. This contributes to the development of more robust and efficient path planning systems for planetary exploration, supported by foundations focused on AI, autonomous systems, and space exploration. The research acknowledges contributions from other researchers in areas like rapidly-exploring random trees and motion planning, emphasizing the need for algorithms that can handle challenging terrains and limited visibility, as demonstrated by recent robotic missions.

Mars and Moon Rover Path Planning Benchmarks

Scientists have created two large benchmark datasets, MarsPlanBench and MoonPlanBench, derived from high-resolution digital terrain images of Mars and the Moon, to address the limited availability of standardized resources for planetary rover path planning research. These datasets, comprising over 2000 2D occupancy maps for Mars and over 36 for the Moon’s north and south poles, represent a significant step towards standardized evaluation of navigation algorithms in realistic planetary environments. The team processed data from the Mars Reconnaissance Orbiter and NASA’s Lunar Reconnaissance Orbiter, specifically utilizing HiRISE and Lunar Orbiter Laser Altimeter data. For MarsPlanBench, researchers generated 2386 grid maps, differentiating based on slope thresholds for defining non-traversable terrain, resulting in maps averaging 200×400 resolution.

MoonPlanBench leverages 36 planar occupancy grids derived from the Moon’s north and south pole landing sites, with a spatial resolution of up to 5 meters per pixel. A publicly available codebase was adapted to facilitate the generation of higher resolution maps and ensure reproducibility. Experiments employed both classical and learning-based path planning algorithms, evaluated on these new datasets and existing benchmarks, to provide comprehensive insights into performance characteristics on planetary terrains.

Mars and Moon Datasets Advance Rover Planning

Scientists have created new benchmark datasets, MarsPlanBench and MoonPlanBench, derived from high-resolution terrain images of Mars and the Moon, to advance research into autonomous rover path planning. These datasets, comprising over 2000 2D occupancy maps for Mars and over 36 for the Moon’s north and south poles, address a critical gap in available resources for testing and refining rover autonomy. Experiments demonstrate that classical path planning algorithms achieve high success rates on the challenging terrains of the Moon’s north and south poles, validating their continued use in missions like those undertaken by NASA. Measurements reveal a clear performance distinction between algorithmic approaches, with classical methods outperforming learned models in terms of success rates, path lengths, and planning times. The team rigorously evaluated representative classical and learning-based algorithms, providing new insights into their capabilities on planetary terrains. This research delivers a foundational resource for the robotics community, with both the datasets and associated code made publicly available to foster further innovation in autonomous rover navigation and exploration, and pre-processing code enables the generation of standardized datasets from planetary digital terrain models.

Mars and Moon Rover Path Planning Benchmarks

This research presents advances in planetary rover path planning, introducing the first large planar benchmark datasets, MarsPlanBench and MoonPlanBench. Derived from high-resolution terrain imagery of Mars and the Moon, these datasets address a critical need for standardized resources to facilitate and evaluate autonomous navigation algorithms. Through rigorous testing of both classical and learning-based path planning methods, the team demonstrates the continued effectiveness of traditional algorithms, such as Dijkstra’s algorithm, in challenging planetary terrains. Results indicate these established methods consistently achieve high success rates in global path planning, a finding corroborated by their current use in the autonomous driving system of the Mars rover Perseverance. The study also sheds light on the limitations of current learning-based models, which struggle to generalize to the complexities of planetary surfaces. This work establishes a foundation for future research by providing a publicly available codebase and datasets, enabling expansion with data from upcoming lunar and Martian missions, and potential investigations into three-dimensional terrain representations and cooperative exploration strategies.