Google open-sources SayCan for language-guided robot planning

Google's robotics team released an open-source version of SayCan for a simulated tabletop environment. In tests on 101 instructions with an Everyday Robots robot, PaLM-SayCan chose the correct plan 84 per cent of the time and executed it successfully 74 per cent of the time, ahead of a version using the FLAN model. The system used a chain-of-thought prompt with 17 worked examples, but it struggled with negated requests such as asking for something that is not an apple. Opening up the code helped spread the idea of using language models as robot planners across the research community.

Source: InfoQ