Korean researchers teach a humanoid robot to walk and handle objects in one sequence without full demonstrations

Researchers at Seoul National University in South Korea have built an AI system that lets a humanoid robot walk to an object, pick it up, carry it and set it down in one sequence, even though it learned walking and grabbing separately, the university said.

The problem

Robots often learn by copying human demonstrations, recorded while a person remotely steers the robot.  Recording a long task that mixes walking and handling objects is hard:  if any step fails, the whole recording is wasted.  Recording walking alone, or grabbing alone, is much easier, but then the robot does not know when to stop one and start the other.

How RISE works

The team's system, called RISE, is a vision-language-action model, a type of AI that looks through cameras, reads a plain-language instruction such as "pick up the cup and put it on the table," and turns that into movements.  Inside it are several "experts," smaller parts that naturally specialize in walking or in handling objects, plus a controller that decides how much to rely on each.  When that mix shifts, the robot treats it as a sign that one step is done and moves to the next.

In computer simulations, RISE completed tasks 87.3% of the time, compared with 73.6% for a version built on GR00T, a robot AI model released by Santa Clara-based Nvidia.  On a real Unitree G1 humanoid, it succeeded 50% of the time versus 30% for the comparison system, with its lead growing on longer tasks.

they must be able to reliably learn behaviors with different characteristics, such as locomotion and object manipulation

The research, led by Oh, was accepted to IROS 2026, a major robotics conference.  The team said it could first help robots move goods in warehouses and factories.

Source:  Seoul National University via TechXplore; RISE project paper.