The depth and detail in this paper (8 page PDF) are daunting, but the gist, as summarized in this post, is straightforward: provided only with observations of a human cleaning (" using a so-called instrumented tool, which is a standard tool equipped with additional sensors") the front of an ordinary sink, a robot arm learn to clean entire sinks with varying surface topology. "The model implicitly learned the correct
tool orientation and position w.r.t. the freeform 3D surface." Michael Franco speculates, "a fleet of robots could learn the basic moves from each other through what's known as 'federated learning' and then they could apply those moves to their individual, specified tasks."
Today: 52 Total: 236 [Direct link] [Share]
