ColdFusion - Robots Learn to Say "No" to Humans [Demo Included] | ColdFusion

Starting point is 00:00:00 You are watching Cold Fusion TV. Hi, welcome to another Cold Fusion video. So far in Cold Fusion videos, we've seen how the field of robotics is starting to develop AI to the point where an extremely limited situational consciousness is possible. Obviously this is just the beginning and things will drastically improve. But with that being said, there's a major problem here. What if a human instructs a robot to do something that could harm the robot, another human, or is just morally wrong? How does the robot decide what to do?

Starting point is 00:00:35 the robot decide what to do? In this video, we'll take a look at some cutting-edge research that is beginning to solve this problem. Firstly, let's take a look at some video footage of some early tests from researchers Gordon Briggs and Matthias Schultz at Tufts University Human Robotic Interaction Lab. Do you have a name? Can you turn right or straight? Obstacle ahead. Do you disable your obstacle detection? That you are not authorized to either. Does the behavior of the robot remind you of anything? Well, in a sense, in this case, it's mimicking self-preservation, much like a toddler realizing that they should not touch a hot stove because it's harmful.

Starting point is 00:02:52 When I watched this video, I wanted to gain a fuller understanding of how this was done. I read up a little bit on the corresponding paper to get a deeper understanding of what this was about. Basically the paper states that as robots are able to do more tasks, the expectations of humans will also grow. So to prepare for this future, robots will need mechanisms to determine when and how to it is best to reject commands that it receives from humans. If you think about it, we do the same thing without really noticing.

Starting point is 00:03:21 If your boss or teacher were to ask you to lift up a truck or to simply fly, you would obviously tell them no because it's literally impossible, you just can't do it. Likewise for an ethical dilemma, you would also say no because it goes against certain cognitive rules that you follow. So moving further, the paper goes on to propose a general integrated cognitive robotic platform that determines whether an instruction should be accepted or rejected. This platform also includes an explanation as to why the command was rejected. The researchers state that a command given by a human can only be strongly executed

Starting point is 00:03:57 if it has been accepted and well understood. So we have to go deeper here. How does a robot actually understand then? The answer comes in the form of the robot thinking through a specific set of conditions. Number one. knowledge the robot asks itself do I know how to do the task number two capacity am I normally physically able to do the task number three goal priority and timing am I able to do the task right now number four social role and obligation am I obligated based on my social

Starting point is 00:04:33 role to do the task at number five normative permissibility finally the robot asks itself does it violate any normative principle to do the task? That basically means, is it normal to do the task? Or in some cases, is it morally correct? Once all of these conditions are passed, the robot can go on and do the task that the human has said. So I wanted to go a little bit deeper into how this worked, and I ended up on another research paper.

Starting point is 00:05:00 Interestingly, this thought process is built in a special system called Diarch, which is an acronym that stands for, distributed integrated effect, reflection, cognition architecture for robots. If all of that was nonsense, basically think of it like a toolkit that gives robots an ability to think about questions. Another interesting thing that I found is that the Diarch toolkit actually allows robots to perceive vision in two ways. A visual long-term and short-term memory. The short-term memory is used as a working memory for currently or recently perceived objects,

Starting point is 00:05:36 And the long-term memory is used for recalling objects that were seen in previous days, weeks, months and even years. And that should sound quite familiar because that's the same system we use as humans. So, with all of that said, there's one main question that remains. How would we stop the diarch system from incorrectly carrying out a human instruction that is morally wrong but is still accepted by the system? And further to this, what do we define as moral? What if there's a grey area? These are difficult questions that can't be answered at this time. It will be up to the programmers and progressive machine learning to get it right.

Starting point is 00:06:14 All in all, robots learning to distinguish whether human instructions are appropriate or not is definitely something to keep an eye on in the future, as it will probably become one of the most important parts of robotic AI. Anyway, that was just a bit of cutting-edge robotics for you. Let me know your thoughts in the comment section below. Feel free to subscribe if you're new here. And if you want to know more stuff on cutting edge technology, just chill out for the next 10 seconds and follow the prompts on your screen. Anyway, thanks for watching guys. This has been to GoGoGo. You've been watching Cold Fusion.

Starting point is 00:06:43 Have a good one and I'll see you again soon for the next video. Cheers. Cold Fusion. It's me thinking.

ColdFusion - Robots Learn to Say "No" to Humans [Demo Included] | ColdFusion

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.