Google decided to combine what is called business with pleasure – as part of a project to study the methods of teaching robots, the company decided in test mode to introduce them into its office and make them assistant waiters. Such robots help live employees with cleaning and bring them goodies. It is reported by vc.ru.
In order for mechanized assistants to be able to recognize not only a specific task (“Bring a banana”), but also cope with complex and abstract questions (“I just ran, I want to replenish energy” or “I spilled porridge, can you help?”), Their creators used neural network learning language models. True, they had to be modified.
Neural networks correctly understood the request from the very beginning, but offered non-working solutions: for example, GPT-3 wanted to vacuum up the spilled liquid, and FLAN only apologized. But the results improved when the engineers combined the PaLM model with robot-specific software to create the PaLM-SayCan system.
Among other things, robots analyze the environment (kitchen), detect familiar objects and learn from experience. Already at the first stages of the experiment, engineers recorded the correct sequence of actions in 84% of cases and the successful completion of the tasks in 74%.
Source: Trash Box