Tuesday, February 9, 2021
Second AI and Data Science Workshop for Earth and Space Sciences (JPL/virtual)
We use geomorphology as the setting, to present insights connecting Bayesian optimization for sequential experimental design, with semantic information acquired using robotics and AI. A key novelty in our approach is the use of physics engines commonly used in robotics and computer gaming, to run ensemble dynamical simulations on rock particles extracted from 3D maps reconstructed using aerial imagery. Our mapping pipeline for geomorphological analysis uses structure from motion (SfM) and deep learning on close-range aerial imagery to estimate spatial distributions of rock traits (diameter and orientation), along sites such as a fault scarp. The pipeline leverages drone-based imagery to help scientists gain a better understanding of surface processes, through estimation of georeferenced orthorectified maps at 2 cm/pixel ground sampling resolution. Expert annotation on a set of image products, are used to train deep neural networks to detect and segment individual rocks spanning the whole site. In effect, the pipeline automatically extracts semantic information (rock boundaries) on large volumes of unlabeled high-resolution aerial imagery. Our prior work in precision agriculture explored the use of robotic imaging to estimate semantic maps of fruits and trunks in farms. Transitioning from mapping to dynamics, we are exploring simulation of rock motion such as tipping and tumbling on digital surface models of a geologic site. Ensemble simulations using Monte Carlo methods will enable better understanding of surface processes such as earthquakes, from observed field data. To simulate rock particles, we are leveraging the Gazebo simulator running the Open Dynamics Engine. We insert reconstructed 3D rock models in the simulator, and experiment with rock motion by displacing the ground, and record rock trajectories. A preliminary example is available at https://youtu.be/M9htQakJEW8 . This combination of geomorphological analysis and rigid body dynamics simulation of particles, can enable information collection in subsequent field missions, efficiently. Tying it all together, our collaborative swarm simulation testbed, OpenUAV, enables end-to-end software in the loop (SITL) simulated missions at test sites, with drones with a variety of actuator and sensor configurations. Our synergistic approach to close the loop on data-driven science has the potential to scale up a variety of earth and space missions.