Deep Learning for Robotic Exploration
Major Professor: Dr. Geoffrey Hollinger
Emerging applications for robotic data collection include ocean monitoring, emergency response and urban search and rescue. At the core of these applications is a robot's ability to make informed decisions on incomplete data. This dissertation addresses this problem by developing novel techniques for modeling and estimating structured environments using deep learning. The proposed methods improve the efficiency of robotics systems across a wide array of applications and scenarios. A challenging problem in robotics is to predict future observations based on previously-recorded data. Robots often operate in built environments that tend to contain some underlying structure, such that newly-visited locations may appear broadly similar to previously visited locations but differ in individual details. The proposed technique exploits the inherent structure of the environment to train a convolutional neural network that is leveraged to facilitate robotic search. We start by investigating environments where the full environmental structure is known, and then we extend the work to unknown environments. Experimental results show the proposed framework provides a reliable method for decreasing the area searched to find a point of interest. We demonstrate the proposed framework increases the search efficiency of a mobile robot in a real-world office environment.
To utilize uncertainty in our decision making and account for dynamic environments, we propose a convolutional LSTM network with bootstrapped confidence bounds as a method for modeling spatio-temporal data. By providing estimates with confidence bounds that are accurate far into the future, multi-step planners can be utilized to improve performance on information gathering missions. This technique is compared to existing environment modeling techniques. We demonstrate that our proposed approach constructs long-horizon estimates with greater accuracy. We also achieve more accurate and more conservative confidence bounds. Validation through simulation shows our technique increases path planning performance in environmental information gathering missions. Robots often require a model of their environment to make informed decisions. In unknown environments, the ability to infer the value of a data field from a limited number of samples is essential to many robotics applications.
In this dissertation, we propose a neural network architecture to model these spatially correlated data fields based on a limited number of spatially continuous samples. Additionally, we provide a method based on biased loss functions to suggest future areas of exploration to minimize reconstruction error. We run simulated robotic information gathering trials on both the MNIST hand written digits dataset and a Regional Ocean Modeling System (ROMS) ocean dataset for ocean monitoring. Our method outperforms Gaussian process regression in both environments for modeling the data field and action selection.
Tuesday, June 4, 2019 at 10:00am
Rogers Hall, 226
2000 SW Monroe Avenue, Corvallis, OR 97331