KAIST's Humanoid v0.9 pushes 35 kg and lifts 20 kg in demo

In a new demonstration video, a humanoid robot moves among a set of objects, shoves some of them across the floor, lifts others, and arranges them into the letters K, A, I, S, T. The robot is Humanoid v0.9, built by researchers at the Korea Advanced Institute of Science & Technology (KAIST), and the clip released on October 19 shows it handling loads of up to 35 kg (77.2 pounds) when pushing and up to 20 kg (44.1 pounds) when lifting.

The difficulty is physical before it is software. When a humanoid pushes a heavy object, it has to keep its balance while applying force. When it lifts one, it has to bear the object's weight and control its posture through the whole movement. KAIST's Dynamic Robot Control & Design Laboratory (DRCD Lab) works on this coordination, in collaboration with RAILAB, which developed the robot's arms. The team's approach does not treat walking and manipulation as separate tasks: the robot moves its body while its arms apply force. The earlier v0.7 platform had shown fast running and Michael Jackson's Moonwalk. Version 0.9 shifts the focus toward physical interaction.

The robot's skills were developed with deep reinforcement learning, a method in which a control policy is trained through repeated attempts and performance feedback, and with learning from human demonstrations, in which examples guide how a task is performed. Onboard visual feedback tells the robot where the objects are and how they sit, so it can adjust its movements as it handles them. The team describes this as a combination of perception and physical interaction, which is what the letter-arranging task was designed to show.

The demonstration is narrower than the headline figures suggest. Pushing and lifting are different tests: the force needed to slide an object depends on friction and on the surface it rests on, while lifting means carrying its full weight, so the two numbers cannot be added up or compared as one payload rating. The video establishes the loads handled in these specific tasks. It does not establish the robot's maximum sustained lifting capacity, nor how it performs under other operating conditions. The training setup, the division of work between the learning methods, and the control architecture are not described, and the video does not show that the same tasks succeed reliably across different objects, surfaces and environments.

So what changes, and for whom? For now, the change is in the research agenda rather than in any deployed machine. Teams building humanoids get a clear marker that whole-body movement and force-based manipulation are being developed together, and the next measurable question is repeatability: whether the same pushes and lifts work across varied objects, floors and conditions. The source does not say what applications the platform is aimed at or when it could be used outside the lab. Until those results are published, Humanoid v0.9 is a research platform that has shown what it can do in a set of controlled tasks.