CAREER: Robust and Autonomous Robot Adaptation in Novel Scenarios

Stanford University, Stanford CA

Investigators

Abstract

A major current technological challenge is to enable deployment of robots into open real-world environments to help advance human welfare. This Faculty Early Career Development (CAREER) project will serve this goal while promoting scientific progress by studying how robots can adapt to new and unfamiliar surroundings encountered during deployment. Adaptation is needed for handling open-world environments since it is remarkably challenging to foresee all the possible scenarios that a robot may encounter. Adaptation demands a degree of autonomy and robustness. This project will study how robots can autonomously adapt during deployment, identify when and how to ask a human for help, and avoid catastrophic failures or getting perpetually stuck in place. The outcomes of this project are expected to have the potential to significantly expand the set of practical applications of robotics, including in manufacturing settings when there is variability in parts and desired configurations, and in the service industry, such as in hospitals and homes where the environment changes frequently based on people’s behavior. This project will also support the development of freely available course lectures and course content pertaining to robotics and machine learning, as well as a mentoring program for undergraduates from groups that are underrepresented in STEM. The central objective of this project is to advance the capability of robots to adapt online during deployment. Robotic reinforcement learning systems can in principle be applied to enable online adaptation, but in practice they are currently ill-equipped to do so. This is because they require supervision and environment resets that are unavailable during deployment. These reinforcement learning systems also do not provide means for robots to identify failures and proactively request interventions in novel environments. This research will develop capabilities that address these challenges and integrate them into a single real robotic system. The research will advance our understanding of: (1) how robots can prepare for unknown situations; (2) how autonomy affects the performance of robotic learning systems; (3) how robots can detect and avoid failures and irreversible states even in new environments; and (4) when automated robot systems should seek external forms of supervision. The developed framework will be thoroughly evaluated on physical robot arms, testing the ability to adapt to a wide variety of unseen circumstances, including new object poses, object materials, and object shapes. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.