Google makes robots smarter, explaining to them what they can and can't do

Google makes robots smarter, explaining to them what they can and can't do

Those who use intellectual voice assistants such as Alice, Siri, etc. may have noticed that technology is getting smarter every day, but the gap between voice control technologies and their implementation in autonomous robotics is still huge, and there are a number of reasons for that.

It is more difficult to teach the robot to perform repeated tasks in controlled spaces without the presence of people, although it is not the easiest task to do. It is much more difficult to teach the robot to do different tasks on the basis of voice commands in spaces where people are also present. It is not about models such as robot vacuumers, which are simply programmed not to touch any objects on the floor.

Google has made some progress in robotic understanding of the natural language that humans can use. Through its natural language processing system, Pathways Language Model has been able to accurately process phrases and understand what a robot really wants, not literally do.

The next challenge is to understand what a robot actually can do. A robot can understand a request to get an object out of a shelf, but the problem is that it can't reach it because the shelf is too high. Google calls "opportunities" what a robot can do more or less successfully. It can be simple tasks as well as complex, multi-stage actions that require a robot to understand its own abilities and the world around it. For example, "In the latter case, the robot will have to break the task into a number of stages — to determine where the liquid is spilled, to go to the kitchen, to find the sponge, to go back, to collect water, to go back to the kitchen to squeeze the sponge, etc. Although it may need to be determined — it may be better to bring the cola can first and then to remove the puddle?

Another problem that robotics face is that language models are not tied to the physical world. For example, on a request, "I spilled my drink, can you help?" the GPT-3 language model answers, "You can try using vacuum cleaners." And that makes sense to her, because the language model associates the vacuumer with the cleaning process. Although the vacuum cleaner is not designed to eliminate the puddle and an attempt to do so could cause it to break down.

As Google says, it's important to teach robots what they can and can't do and what it makes sense to do first and foremost in different situations.