Researchers at the Swiss École Polytechnique Fédérale de Lausanne have developed Kinematic Intelligence, a framework that enables robots to learn skills from a single human demonstration and transfer them to different hardware without retraining. The AI-free system avoids joint singularities, ensuring safe operation across varied robot designs. The work is detailed in a paper published in Science Robotics.
Roboticists have long faced challenges in transferring learned skills between robots with different designs, such as varying joint lengths or orientations. A team led by Sthithpragya Gupta at EPFL addressed this with Kinematic Intelligence, which embeds a robot's physical constraints—like joint limits and singularities—directly into its control policy from the start. 'With new designs come different capabilities and constraints,' said co-author Durgesh Haribhau Salunkhe. 'The problem is to adapt to these constraints and capabilities—to faithfully replicate the actions demonstrated by a human.'
