Researchers led by Baicheng Li have built a training method that lets a robot carry out a skill that was only ever demonstrated on a different robot, according to a paper on arXiv. Robot demonstrations come from many different machines, so many pairs of robot and skill never appear in the data.

The key idea is that a demonstrated command gives a target position for the tool of every robot that can reach it. Calibrated kinematics turns that target into training signals for the other robots, without their joint data for that skill, and reachability checks pick which targets to use. A visual policy predicts shared actions, which each robot's own decoder turns into motion.

Repeated training and tests with three model backbones in two simulators showed better transfer to pairs of robot and task that were missing from the data. The abstract gives no figures.

Source: arXiv Robotics (cs.RO)

Background

Baicheng Li

Robotics researcher

Robotics researcher.