Robotic exploration is a desirable goal for small-scale SWARM systems. From practical applications like rendezvous to pursuit evasion, the ability to map out an environment relative to one's own position is of vital importance to SWARMs. Simultaneous Localization and Mapping (SLAM) techniques are well suited to this problem. While traditional SLAM methods like Visual Odometry with the FAST interest point detector generally perform efficiently and well at these tasks, they are insufficient when dealing with the levels of noise we expect to encounter with low-resolution, millimeter-scale, grayscale cameras. Our approach seeks to address this problem through a combination of Control Feedback and Unsupervised Learning techniques. Benchmarks on the KITTI Odometry dataset shows significant gains in settings where images are grossly corrupted by Gaussian noise. Control Feedback techniques alone can provide similar performance to the noiseless setting in these situations; Unsupervised Learning techniques provide similar and sometimes even better performance than FAST while performing fewer instructions in the worst case. Through these techniques, we aim to bring robotic exploration at the scales we would expect for a SWARM system closer to reality.




Download Full History