Researchers led by Tianqi Gong have built GTI, which helps vision-language-action models keep the overall goal of a long manipulation task in view, according to a paper on arXiv. The method uses the current view and the history of the task to work out which subtask is under way and how far it has got.
For each subtask it generates two target images, one of the moment of contact and one of the finished step, and these, with a short instruction, steer the model. When the robot keeps drifting off course, GTI rolls back, looks again and re-evaluates the phase.
With the GR00T model, success on seen composite tasks in RoboCasa365 rose from 14.6% to 31.8%. In three tasks on a real PiPER robot arm, success on the full task rose from 21.7% to 59.2% on average.



