Google's PaLM-SayCan lets a large language model plan for a real robot

Google described PaLM-SayCan as the first use of a large-scale language model to plan for a real robot, pairing its PaLM model with robot skills on an Everyday Robots helper. The language model suggests useful steps while learned value functions check which are feasible, so that for a spilled drink the robot fetches a sponge rather than a vacuum. Compared with a weaker model, planning success improved by 14 per cent, with the largest gains on long tasks of eight or more steps. The robots were being tested fetching snacks in Google's office kitchens. SayCan was an important early step towards robots that understand everyday spoken requests.

Source: Google