Towards Improving and Extending Traditional Robot Autonomy with Human Guided Machine Learning

Traditional autonomy among robotic and other artificial agents was accomplished via automated planning methods that found a viable sequence of actions, which, if executed by an agent, would result in the successful completion of the given task(s). However, many tasks that we would like robotic agents to perform involve goals that are complex, poorly-defined, or hard to specify. Furthermore, significant amounts of data or computation are required for agents to reach reasonable performance. As a result, autonomous systems still rely on human operators to play a supervisory role to ensure that robotic operations are completed quickly and successfully. The presented work aims to improve the traditional methods of robot autonomy by developing an intuitive means for(human operators to adapt/mold the behaviors and decision making of autonomous agents) autonomous agents to leverage the flexibility and expertise of human end users. Specifically, this work shows the results of three machine learning-based approaches for modifying and extending established robot navigation behaviors and skills through human demonstration. Our first project combines Imitation learning with classical navigation software to achieve long-horizon planning and navigation that follows navigation rules specified by a human user. We show that this method can adapt a robot's navigation behavior to become more like that of a human demonstrator. Moreover, for a minimal amount of demonstration data, we find that this approach outperforms recent baselines in both navigation success rate and trajectory similarity to the demonstrator. In the second project, we introduce a method of communicating complex skills over a short-horizon task. Specifically, we explore using imitation learning to teach a robot the complex skill needed to safely navigate through negative obstacles in simulation. We find that this proposed method could imitate complex navigation behaviors and generalize to novel environments in simulation with minimal demonstration. Furthermore, we find that this method compares favorably to a classical motion planning algorithm which was modified to assign traversal cost based on the terrain slope local to the robot's current pose. Finally, we demonstrate a practical implementation of the second approach in a real-world environment. We show that the proposed method results in a policy that can generalize across differently shaped obstacles and across simulation and reality. Moreover, we show that the proposed method still outperforms the classical motion planning algorithm when tasked to navigate negative obstacles in the real world.