Northwestern's RL system uses physics priors for dexterous robot hands

Researchers at Northwestern University have developed a reinforcement learning approach that embeds physics priors to achieve robust in-hand manipulation without relying on external sensing. The team, led by Yifei Chen, Shihan Lu, Ed Colgate, and Kevin Lynch, addressed the challenge of finger-object contact uncertainties and gravitational disturbances by introducing two complementary priors. The first is a global grasp-quality prior derived from classical grasp analysis, used as a dense reward-shaping term to encourage well-distributed contacts with improved worst-case wrench resistance. The second is a local contact-geometry prior expressed through fingertip curvature, which mechanically shapes the contact interface toward task-aligned rolling while reducing off-axis drift.

The system was evaluated on a multifingered robotic hand manipulating three different objects at four palm orientations. Results showed significant improvements in rotation efficiency, grasp stability, and disturbance rejection, suggesting that integrating physics priors into both the learning algorithm and fingertip morphology enhances task robustness and sim-to-real transfer. This work bridges classical robotics theory with modern RL, offering a practical path toward dexterous manipulation in real-world environments where external sensors are unavailable or unreliable.