Dexterous robot hands (Credit: Midjourney)
In dynamic, unstructured environments like ship decks and even home kitchens, robots today still struggle to perform precision tasks such as tightening bolts or handling wires. This makes critical ship maintenance tasks difficult.
USC researcher, Erdem Bıyık, aims to advance robots’ finger manipulation and integrate human feedback to enable real-time learning for robots in an upcoming three-year, $750,000 project funded by the Office of Naval Research.
As an assistant professor at USC Viterbi School of Engineering’s Thomas Lord Department of Computer Science and USC School of Advanced Computing, Bıyık was selected for the 2026 ONR Young Investigator Program (YIP) Award for his project titled “Learning Dexterous Robot Hand Behaviors from Multimodal Human Feedback.” He also holds a joint faculty appointment at the Ming Hsieh Department of Electrical and Computer Engineering and leads the Learning and Intelligent Robot Assistants Lab (Lira) at USC.
The ONR YIP is a highly competitive program that supports early-career researchers pursuing innovative solutions to challenges faced by the Navy and Marine Corps. Bıyık was one of 23 recipients selected nationwide this year.
Erdem Bıyık
With a background in computer science, robotics and electrical engineering, Bıyık will work to advance dexterous robotic manipulation—enabling machines to perform precise, human-like tasks using multi-fingered hands. While most robots today rely on simple grippers, his research aims to unlock more sophisticated capabilities, such as tool use, fine assembly, and adaptive handling in complex environments.
His work addresses a key limitation of current systems: even a single misplaced finger can cause failure in high-precision tasks, sometimes with damaging consequences. By developing more advanced, multi-fingered hands, Bıyık aims to enable robots to operate hand tools, such as hammers, and perform tasks like tightening knobs and bolts that require a high level of precision. Such advanced finger manipulation capabilities ultimately allows robots to operate in unstructured and dynamic environments, like shifting ship decks and busy kitchens.
Another key part of the project is introducing a new framework for teaching robots through multimodal human feedback.
Because it is traditionally difficult to provide robots with the precise feedback needed to correct minute physical errors, Bıyık explained his “goal is to make robot learning more aligned with how humans teach each other. By integrating multiple forms of feedback, we can enable robots to acquire complex skills faster and more reliably.”
By combining visual demonstration—where robots learn by observing human actions—with natural language guidance, such as real-time verbal corrections, Bıyık’s work seeks to create more intuitive and efficient ways for humans to train robotic systems.
The application and technologies developed through this project goes beyond naval applications, and can be extended to everyday environments, supporting tasks in homes and even warehouses or service industries that require high levels of precision and adaptability.
By developing robots that can learn directly from human input and operate with greater dexterity, Bıyık’s work aims to bridge the gap between controlled laboratory systems and real-world deployment.
Bıyık also added that Lira lab plans to release open-source tools and datasets, helping accelerate progress in robotic learning across the broader research community.
Published on April 7th, 2026
Last updated on April 7th, 2026