Robots can now learn and share skills even with different hardware thanks to a new framework developed by researchers at EPFL. Dubbed Kinematic Intelligence, this system addresses a major challenge in robotics where learned tasks are typically tied to specific machines.
Traditional methods of teaching robots involve demonstrating tasks, but variations in robot design-such as link lengths or joint orientations-cause these learned behaviors to fail on new machines. This often necessitates retraining from scratch.
EPFL's solution embeds a deep mathematical awareness of a robot's physical limitations directly into its control policy. The system maps out singularities, mathematical danger zones where robot motion can become unstable, and joint limits. By classifying robots into categories based on these limitations, Kinematic Intelligence allows robots to navigate around singularities safely.
This AI-free framework was tested on three distinct robotic arms: a Duatic DynaArm, a KUKA LWR IIWA 7, and a Neura Robotics Maira M. A single human demonstration of a complex task sequence was successfully transferred and executed by all three robots, even when their roles and positions were shuffled without retraining.
While Kinematic Intelligence ensures mechanically safe motion, future advancements will focus on integrating advanced sensing and context-sensitive decision-making for unpredictable environments and sensitive applications like medicine.