Robust Failure Recovery for Robotic Manipulation

This project will fund research that looks to develop techniques to help robots recover from grasping failures in various settings, including homes and underwater environments. Practical applications include picking up objects from cluttered tables, clearing disaster debris, and rescuing marine life trapped in abandoned fishing gear. These advancements will benefit multiple sectors of the US economy, from fisheries to household consumer goods.

As industries increasingly rely on robotic systems for tasks in outdoor spaces, crowded areas, underwater environments, and other unpredictable settings, the ability to grasp moving or cluttered objects becomes essential. Failures are inevitable, so effective recovery strategies are crucial to improving the overall success rate of robotic grasping. However, current techniques often struggle to adapt after a failure occurs, limiting their reliability.

Beyond research, the educational impact of this project includes several activities such as providing undergraduate students at Oregon State University with the opportunity to gain hands-on experience in robotic grasping and introducing K-12 students to search-based planning concepts by collaborating with local school robotics clubs in Oregon. Additionally, the project team will demonstrate grasping algorithms on a physical robot to enhance the daily experiences of residents with dementia at the Grace Center for Adult Day Services in Corvallis, Oregon.

This grant aims to provide fundamental algorithmic contributions to heuristic search-based planning under uncertainty, specifically for addressing grasping failures caused by different factors, primarily: (1) unknown object parameters such as mass and friction coefficients, (2) unknown dynamics of a moving object, and (3) unknown scene segmentation and interaction dynamics in a cluttered environment. The goal is to improve the grasping success rate and reduce the number of attempts required to achieve a successful grasp while ensuring real-time planning speed and theoretical guarantees.

The core premise of this work is to compute policies, rather than plans, for failure recovery. The key insight is to develop conditional policies that account for anticipated failure outcomes of a grasp, where each outcome explicitly minimizes the probability of failure for subsequent attempts. The PI will involve learning dynamic Bayesian networks, deep neural networks, and hybrid models to enhance generalizability.

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.