A person and a robot lift a car tyre together, moving between low walls. Moments later, they carry a long wooden plank. Weight is not the only challenge: when the person turns, stops or bends their knees, the robot must keep cooperating.
Researchers at Carnegie Mellon University and the University of Texas at Arlington are training that cooperation with HALO. They use a Unitree G1, a humanoid robot made by a Chinese manufacturer. The research focuses on moving together with someone who does not always follow the same pattern.
Carrying together keeps changing
Anyone who has carried a sofa through a doorway with another person knows the problem. You need to turn together, adjust your steps and make sure your partner does not lose their end. A robot that simply follows its own route could disrupt the cooperation.
The experiments on the research website make this visible. A person and robot move an object through an opening, transport a load in a confined space and manoeuvre with a long plank. In another test, the robot adjusts its posture when its partner lowers their body.
Practising in a computer first
The team trains in the Isaac Lab simulation environment. Different digital partners learn to complete a task together there. This provides more variation than practising with a single predetermined movement pattern.
HALO intervenes during that learning process. The method corrects updates that would cause the partners’ cooperation to break down. The aim is for the control system to learn to take the other carrier’s actions into account.
In the real experiments, a motion-capture system helps record positions. That is an important step between practising in a computer and having a robot that can work with people anywhere, without this kind of laboratory setup.
A future colleague must be able to wait
The researchers also test what happens when the person stops unexpectedly. The robot must then adjust its movement rather than keep pulling the load. These small interruptions are precisely what makes working together different from repeating a trick.
The research results also show that this is not yet a flawless helper. The real-world tests involved five repetitions per measurement; with the long plank, the load was dropped in some trials even with HALO. The footage offers a glimpse of robots working alongside people. For everyday use, that cooperation must also become reliable beyond the test setup.
