Ajna was highlighted on the cover of Science Robotics journal in Aug 2023. The paper can be found here for free. This work is a collaboration between the University of Maryland, College Park (PRG-UMD) and Worcester Polytechnic Institute (PeAR-WPI). Please refer to the wiki for instructions on running the code and some research tips and tricks.
Please check out the video for a description of the approach.
We present an approach to estimate uncertainty of any neural network with a loss function that can be adapted to modify any existing loss function for any class of robotics problems. We showcase the approach's efficacy in flying through gaps, navigating through a static environment, dodging dynamic obstacles and segmenting objects from a scene.
If you find our work useful please do cite us as follows:
@article{
doi:10.1126/scirobotics.add5139,
author = {Nitin J. Sanket and Chahat Deep Singh and Cornelia Fermüller and Yiannis Aloimonos },
title = {Ajna: Generalized deep uncertainty for minimal perception on parsimonious robots},
journal = {Science Robotics},
volume = {8},
number = {81},
pages = {eadd5139},
year = {2023},
doi = {10.1126/scirobotics.add5139},
URL = {https://www.science.org/doi/abs/10.1126/scirobotics.add5139},
eprint = {https://www.science.org/doi/pdf/10.1126/scirobotics.add5139},
}
Copyright (c) 2023 Perception and Robotics Group (PRG-UMD) and Perception and Autonomous Robotics Group (PeAR-WPI)