Some robot AI works by imagining the future: it pictures how the scene will look next, then picks its moves to match. That foresight helps, but redrawing the whole picture every time the robot updates its plan is slow, so the robot reacts late. Researchers from the University of Southern California, Brown University, Fudan University and Toyota Research Institute say their method, Rolling-WAM, keeps the foresight and cuts much of the wait. Faster updates help any robot that has to keep correcting itself as things change around it.
The team compares the trick to how people plan. Think of driving: you know exactly how you will steer in the next second, while a turn a mile ahead stays a rough idea until you get closer. Rolling-WAM works the same way, finishing only its very next move in full detail and keeping later moves as rough sketches that sharpen over time. After each move it takes a fresh look through its camera, keeps the sketches, adds a new one at the end and carries on, so it does not have to start over from scratch each time.
The researchers report that, in a simulated setup, this makes its plan updates more than four times faster than the standard version of this kind of system. They tested it on a Unitree G1 humanoid robot that had to put a plush dog in a box, stack plates and pour beads from a bottle into a glass, and a try counted only if the robot finished without human help. On average it succeeded more often than the other robot systems it was compared with, and the team says it moved more smoothly, with fewer of the pauses that sometimes stalled the standard version. For teams building robots that plan by imagining, it could offer a way to react faster without giving up that foresight.
The team says its code and trained models are being prepared and will be released soon. The researchers also name a limit: when a scene changes very quickly, the robot's saved sketches of the future can fall behind and steer it wrong, especially when it plans far ahead. They point to adjusting how far ahead the robot looks on the fly, and letting it move and think at the same time, as ways to make it more responsive.
In the paper, Rolling-WAM was tested in simulated scenes and on a single humanoid robot, and its speed figures come from the team's own controlled setup. How well it copes with fast-changing scenes outside those tests is still an open question.