A new neural network input module improves humanoid locomotion (news.engin.umich.edu)

Researchers at the University of Michigan have developed a novel neural network component, PRISM (Polynomial Representations for Interaction-Structured Motor Control), designed to enhance robotic motor control, particularly in humanoid locomotion and object handling. This new approach challenges traditional methods that rely on general-purpose neural networks, which often struggle to effectively manage the complex physical interactions involved in robotic movement. By learning multiplicative relationships among sensor observations, past actions, and task commands, PRISM allows robots to navigate unfamiliar environments without needing larger networks or additional sensors, thereby streamlining the process of physical interaction learning. The significance of PRISM lies in its ability to leverage existing robot observations to draw meaningful conclusions about motion and force relations, leading to more reliable and adaptable movement. Unlike conventional multilayer perceptrons that learn interactions indirectly, PRISM directly incorporates these multiplicative relationships, allowing for faster and more efficient learning without the need to manually define physical equations. The component can be seamlessly integrated into current robot learning systems without altering their training objectives or sensors, marking a substantial advancement in the field of AI and machine learning for robotics. This breakthrough was presented at the 10th Conference on Robot Learning (CoRL 2026) and is poised to improve how robots can operate in dynamic environments.