The National Science Foundation has awarded a National Robotics Initiative grant to CAIRO Lab based on our proposal, “Life-Long Learning for Motion Planning by Robots in Human Populated Environments”, to fund research that will help create robots capable of teaching themselves to socially navigate safely amongst people.
Long-term, safe interaction in human-populated environments is essential to the widespread adoption and continued deployment of robotics. Accordingly, an important aspect of behaving safely is the ability to use one's observations and past experiences to set realistic expectations about others' behaviors, operating in such a way that minimizes the risk of failure or causing harm. Just as humans are capable of adapting their behavior to changing circumstances and understanding possible failures, robots must similarly understand risks and adapt their behaviors to safely complete tasks in densely populated or dynamic environments. To enable these capabilities, this project will develop methods to provide mobile robots with the ability to build their own models of activity-based risk and human behaviors, using them to mitigate dangerous or risk-prone scenarios in the presence of uncertainty. These methods are broadly applicable across service robots, manufacturing robots, and close-quarters collaborative robots, enabling the effective and safe integration of such systems into human environments that have otherwise been inaccessible for enhancement or improvement through robotic automation. By contributing toward making such applications possible, results from this research will benefit the U.S. economy and society at large, while simultaneously contributing to the competitiveness of U.S. manufacturing on grounds of safety and efficiency through advanced and flexible automation. This work involves contributions from many complementary areas of research, including robotics, human factors, manufacturing, and computer science through both academic and industry collaborations.
To address challenges in safe, life-long learning as applied to long-term robot operation, this research focuses on the development of novel algorithms for self-supervised human-aware motion planning and context-aware trajectory optimization. Advancing the state-of-the-art in self-supervised learning through self-determined curriculum development and learning through experiences collected during live deployments, the work will additionally develop new transfer learning techniques to generalize experiences and behavioral models across contexts. Robots incorporating these research products will be able to learn to anticipate and mitigate risks and potential failure modes during task execution in uncertain environments by leveraging their own self-collected knowledge. This award targets fundamental advances in social navigation and task execution by integrating human motion modeling and robot motion planning with contextual understanding, achieving applicable and real-world grounded results through long-term mobile robot deployments in public spaces.