How one can educate the identical ability to totally different robots

The meeting line activity setup. Credit score: 2026 LASA EPFL CC-BY-SA.

By Celia Luterbacher

In right now’s manufacturing environments, upgrading a robotic fleet typically means ranging from scratch – not solely changing {hardware}, but in addition reprogramming duties. Even when two robots are constructed to carry out related jobs, totally different joint preparations or motion limits imply {that a} activity programmed for one robotic typically can’t be used on one other. Enabling abilities to switch instantly between robots may make these programs extra sustainable and cost-efficient.

To fulfill this problem, researchers within the Studying Algorithms and Methods Laboratory (LASA) in EPFL’s Faculty of Engineering have developed a brand new robotic management framework referred to as Kinematic Intelligence. The tactic takes a human-demonstrated activity, mathematically converts it right into a normal motion technique, after which adapts it in order that totally different robots can carry out it primarily based on their bodily design. The analysis has been printed in Science Robotics.

“This work addresses a long-standing problem in robotics: switch a realized ability throughout robots with totally different mechanical constructions, whereas guaranteeing secure and predictable habits,” says LASA head Aude Billard. “This strategy may considerably cut back the time and experience wanted to deploy robots in real-world settings.”

Kinematic Intelligence for transferable robotic studying

To construct their framework, the researchers first took human-demonstrated object‑manipulation duties – corresponding to putting, pushing and throwing – and recorded them utilizing motion-capture know-how. Then, they mathematically transformed these recorded duties into normal motion methods. Additionally they developed a scientific classification of the bodily limits of various robotic designs, together with how far their joints can transfer and which positions they need to keep away from to stay steady. The framework then makes use of this classification to robotically tailor the final motion methods to totally different robotic our bodies, making certain they’ll perform duties safely inside their mechanical limits.

In an meeting line experiment, a human demonstrated a activity by pushing a picket block off a conveyor belt onto a workbench, putting it on a desk, and at last throwing it right into a basket. By utilizing Kinematic Intelligence, three utterly totally different industrial robots have been capable of reproduce this similar sequence safely and reliably.

“Every robotic dealt with totally different steps of the duty, and the system carried out efficiently even when the step allocation was modified,” explains LASA PhD scholar and co-first creator Sthithpragya Gupta. “Every robotic interprets the identical ability in its personal manner, however at all times inside secure and possible limits.”

In direction of scalable and future-ready robotics

The researchers purpose to increase the framework to settings corresponding to human-robot collaboration and pure language-based interplay. For instance, Kinematic Intelligence may enable an individual to instruct a robotic with easy instructions at dwelling, without having for technical programming. The strategy can be related for rising robotic platforms, the place speedy {hardware} evolution implies that right now’s machines might quickly get replaced by newer variations. Enabling seamless switch of abilities throughout such platforms may play a key function in making them sensible and scalable.

“Our purpose is to take away the necessity for technical experience whereas nonetheless making certain secure and dependable operation,” summarizes LASA scientist and co-first creator Durgesh Haribhau Salunkhe. “The consumer brings the concept and the specified habits, and the robotic ought to maintain the remaining.”

Reference

Display as soon as, execute on many: Kinematic intelligence for cross-robot ability switch, S Gupta, D H Salunkhe, A Billard, Science Robotics (2026).