Researchers led by Licheng Zhong have built Progress Predicts Policy, or PPP, which controls a robot by learning how far a task has come instead of learning the actions directly, according to a paper on arXiv. The progress is a single number that depends on the scene and the instruction in words.

The robot's commands come from how this number changes with the action: for continuous commands from its gradient, scaled by a learned gain, and for discrete ones from the differences between options. This needs neither a separate part that outputs actions nor a search while the robot runs. The timing of the demonstrations sets the progress values, so no extra labels are needed, and the number can also show how far a task has come during the run.

In tests on simulated and real robot manipulation, PPP did better than behavioural cloning with the same architecture. The abstract gives no figures.

Source: arXiv Robotics (cs.RO)

Background

Licheng Zhong

Robotics researcher

Robotics researcher.