Imagine a robot that can do your laundry, make your bed, cook dinner or stock a grocery shelf. Until now, robots have been trained on individual actions, but getting them to handle complex, changing tasks has proven difficult despite huge investment in robotics.
A team at École Polytechnique Fédérale de Lausanne in Switzerland, led in part by robotics scientist Sthithpragya Gupta, reports progress. In a paper published in Science Robotics they describe a machine-learning approach that leverages kinematic intelligence — the robot’s internal model of how its own body can move — to learn from human demonstrations. In demonstration videos, single-arm robots watch a person toss a ball into a container and then reproduce the behavior while adapting to their different positions and nonhuman bodies. The learned motions can also be transferred to other robots.
Gupta says this approach could let robots perform subtler, variable tasks — for example making coffee to someone’s taste — instead of only repeating narrowly programmed routines. A long-standing weakness of robots is brittleness: an action learned in one context can fail when small things change, such as a shifting opponent in sports or different lighting. Translating human adaptability into robot behavior has been hard, and this work aims to close that gap.
Experts call the work a meaningful advance. Robert Platt of Northeastern University described it as significant for future robotics, while cautioning that predicting exact timelines is hard given rapid AI adoption.
The development also provokes philosophical and safety questions. If robots can self-correct and teach others, does that imply self-awareness? AI researcher Susan Schneider warns that learning and self-correction are not the same as consciousness, but she and other safety experts say the growing capabilities raise red flags: more advanced systems could be misused or weaponized.
The researchers have built safety measures into current systems, but they acknowledge the need for stronger guardrails and regulatory frameworks about who can operate such robots and how they are deployed. As robots become more adaptive and transferable, society faces choices about oversight, deployment and risk management. It is an exciting step forward, but one with important uncertainties.