They are like little golden feet
For the most part, current robberies operate with a memory almost like that of a little golden foot. They manage to process images and information for a few seconds, but they quickly lose the context of what they were doing. This greatly limits the type of task that a machine can execute.
Most robos manage to complete quick actions, but longer tasks like organizing a kitchen, preparing food, or performing a sequence of complex steps have always been a challenge, now a company called Physical Intelligence announced a new artificial intelligence architecture that can change that. The technology was dubbed multiscale embedded memory or simply MEM.
The idea is to allow robots to maintain focus on a task for up to 15 minutes, remembering what they have already done and what they still need to do, to achieve this, the system uses two different types of memory, video-based short-term visual memory, this system allows the robot to remember recent details of the environment, for example, it can remember where an object is, even if its own arm momentarily blocks vision, and long-term textual memory, instead of storing all the video frames, the robot creates small summaries in language natural of what happened, something like an internal note, “I opened the refrigerator door,” “I took the plate,” “I cleaned the counter.”
This type of adaptation is essential for real-world tasks.
These small narratives help the robot maintain the context of the task it is executing and that brings an important advantage, the ability to learn from its own mistakes. And a test demonstrated by the researchers, a robot tried to pick up an object and failed on the first attempt, but as it remembered the recent error, it automatically adjusted the strategy on the second attempt and managed to complete the action.
For researchers, this memory can be an important step in bringing robots closer to real working environments, well, if robots are becoming smarter, then another question arises.
