ERI: Redefining Robot Intelligence Through Contact-Level Reflexes and Physical Continual Learning for Fast, Reliable Task Adaptation in Unstructured, Contact-Rich Environments

This NSF ERI project aims to improve the safety and reliability of robots in tasks that involve physical contact with objects. Today’s robots rely heavily on accurate sensing, such as computer vision, which can fail in low-light or cluttered real-world environments. As sensing errors build up, a robot may think it is in the right place when it is slightly off, causing it to move incorrectly, leading to unintended impacts and preventing reliable grasp or proper alignment of parts. This project addresses this gap by enabling robots to sense contact, recognize when it is incorrect, and immediately adjust their motion or grip without waiting for human intervention or stopping the task. The project will bring transformative change by embedding biological reflex-like responses directly at the point of contact via a unified sense-interpret-react framework utilizing tunable compliant hardware, contact interpretation, and rapid local adjustment. The intellectual merit of the project includes developing a new, scalable and generalizable approach to designing embodied robotic systems that integrate sensing, physical interaction, and control to ensure reliable operation under uncertainty. The broader impacts of the project include improving the safety and reliability of robots in repair, manufacturing, infrastructure maintenance, agriculture, and healthcare, while also supporting education and accessibility by training students, releasing open-source tools, and engaging students through outreach activities. Technically, the project introduces a new paradigm for contact-rich robotics based on tunable soft interaction elements that filter contact disturbances and determine when corrective actions should be triggered. The mechanical interface acts as a disturbance filter that senses contact through changes in force and deformation, while a spiking neural network (SNN) interprets these signals to identify contact conditions, such as slip or misalignment, and selects an appropriate reflex from a predefined set of recovery behaviors. Once activated, a hybrid control law computes a corrective subgoal, switches the robot from nominal execution to a reflex mode, and guarantees stable return to the original task using Lyapunov-based methods. The system also incorporates an experience buffer that stores prior interaction outcomes, allowing the robot to improve reflex selection and parameter tuning across different objects and tasks. By combining physical adaptation, learning, and provably stable control, the project provides a unified framework that reduces reliance on centralized computation and enables robust operation in uncertain, contact-rich environments. 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. NSF Award ID: 2552333 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Karishma Patnaik | Institution: Regents of the University of Michigan - Dearborn, Dearborn, MI | Award Amount: $198,792 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2552333 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2552333.html