Deep Learning-Based Lunar Crater Terrain Relative Navigation

A new algorithm that pairs a deep-learning crater detector with an Extended Kalman Filter could help future lunar landers pinpoint their location using only orbital imagery, according to research posted to arXiv [1]. The terrain relative navigation (TRN) system is designed to address the challenge of accurate position estimation for autonomous vehicles landing in environments with sparse terrain features [1]. The detector was built specifically for the NASA Crater Detection Challenge and analyzes crater features from monocular images acquired from orbit [1]. Matches with craters in a global database are identified through a Hungarian assignment approach followed by a consensus-based outlier removal method [1]. The resulting measurements feed an Extended Kalman Filter that estimates spacecraft pose in the Lunar-Centered Lunar-Fixed frame, with altitude aiding information used to constrain radial drift [1]. Simulation results indicate the algorithm can recover from an initial location error of up to 5 km, reducing navigation error to a few hundred meters [1]. The authors note that maintaining crater feature correspondences requires matching image resolution and scene scales to the detector’s training set distribution [1]. The work arrives as NASA and other space agencies invest in technologies for precision autonomous landing. NASA has identified lidar as a key technology for enabling autonomous safe landing of future robotic and crewed lunar vehicles [5]. The Ingenuity helicopter on Mars demonstrated the value of autonomous navigation in off-world settings, completing 72 flights over nearly three years before its navigation system struggled over featureless terrain on its final flight, resulting in a crash landing [4]. The last time humans walked on the Moon was during Apollo 17 in December 1972, a mission that set records including the longest crewed lunar landing at 12 days and 14 hours and the largest lunar-sample return of approximately 115 kg [3]. Those astronauts drove a Lunar Roving Vehicle that achieved a top speed of 11.2 miles per hour on its final mission [11]. Future missions will demand far more autonomy. The proposed Endurance rover concept, identified as the highest priority strategic mission for NASA’s Lunar Discovery and Program, would traverse 2,000 kilometers across the lunar farside to collect samples from the South Pole-Aitken basin [6]. A separate study using Lunar Reconnaissance Orbiter imagery has shown that deep generative visual models can automatically retrieve scientifically interesting surface features, including human landing and spacecraft crash sites, from the petabytes of data the orbiter has returned over the past decade [8]. The crater-based TRN algorithm adds to this growing toolkit of machine-learning approaches for lunar exploration, though its performance depends on careful alignment between orbital imagery characteristics and the detector’s training data [1].