Google has partnered with Everyday Robots, a subsidiary of Alphabet, to create waiter robots that can respond to complex requests. Robots will deliver snacks and drinks to employees.
Related article: best laptops for machine learning in 2022

Most modern robotic systems are only capable of following very short, specific commands, such as "Bring an apple." They find it difficult to complete complex tasks and reason about abstract goals, and the request "I just worked out, bring me something useful for a snack" will confuse them, points out Google Research.
The company has developed the PaLM-SayCan method, which combines AI language models with robot learning algorithms. With this method, the robot acts as the "hands and eyes" of the language model.
The user gives the robot a request, which the language model turns into a sequence of steps to perform. This sequence is filtered using the robot's skill set to help determine the most feasible plan given its current state and environment. The robot interprets the command, compares it with its capabilities and breaks it down into smaller steps.
Drawing up a chain of steps became possible thanks to the introduction of a language neural network trained on Wikipedia, social networks and other sites into robots. Similar AI is at the heart of chatbots and voice assistants, Google says, but it hasn't been applied to robots as widely before.
So far, robots perform several dozen simple actions. When asked to help clean up spilled water, they admit that taking a sponge is the smartest course of action. According to a company blog post, the introduction of more sophisticated language AI increased the success rate of robots in executing commands by 13%.