Google earlier announced the use of artificial intelligence modelsPaLM-SayCan, which will make it easier for assistive robots created by parent company Alphabet to understand spoken commands from users and perform tasks correctly.
Because the semantics of what robots can understand are usually different from what the human brain can comprehend, for example, the commands they can recognize must be complete and precise, and the content they can currently understand cannot be overly complex or abstract. Therefore, they can only execute simple descriptive commands such as "Pick up an apple" but may not understand the actual needs of commands such as "I just worked out, can you get me a healthy snack?"
Even though it's now possible to understand the meaning of human sentences through large language models like GPT-3, there's still a long way to go before robots can actually understand the needs behind the language we use in our daily lives. This is because the human brain has a certain degree of imagination about the context behind the sentences. For example, when someone spills a drink and asks for help, they naturally understand the need to clean up the situation, so they might pick up a rag to wipe the floor or a broom to clean up broken items. However, robots are currently unable to make such associations and may only be able to compare feasible approaches from existing databases, which often leads to irrelevant answers.
The PaLM-SayCan model proposed by Google can further assist robots in determining the underlying meaning behind human sentences, thereby converting relevant sentences into instructions that the robots can correctly execute. Furthermore, by increasing task correspondence and skill feasibility judgment, the accuracy of instruction execution can be enhanced.
Taking the case of spilled drinks as an example, after receiving the keyword "spilled drinks", the robot will further associate all the response methods related to the word combination of "spilled" and "drinks". Finally, combined with judgment bases such as scene recognition, it can conclude that tasks such as "cleaning" and "wiping" are required. Ultimately, the robot will choose to take a sponge to wipe up the spilled drink on the table.
Currently, Google hasGitHubThe PaLM-SayCan model is publicly available for interested developers and design teams to use and test.
