In the past few decades, robots have become more accessible to industries and research. However, we are still far from what the sci-fi movies have presented many years ago: the multi-purpose robot-servants that are able to take over with ease any chore in the house.
Why can’t we have that kind of robots yet? Well, undoubtedly, it’s a multi-step process that involves many different areas to collaborate together. However, what fascinates me the most, as a computer scientist, is the part of autonomous navigation and optimal decision making by the robot.
As the word ‘autonomous’ suggests, there should be no external input but the robot’s ‘brain’. What’s more appropriate for that, than reinforcement learning method. In that kind of training, you provide the barely minimum tools to the robot; how to move around. With no other prior knowledge, you ask the so-called ‘agent’, to understand what’s the optimal way to behave in an unknown environment.
Having an agent able to move around by itself, we want to make the navigation meaningful. We want to store the map of the environment we previously explored, be able to revisit places we’ve seen before, and know at any time – any place where we are relative to the map. The best way to do all that is called SLAM (Simultaneous Localization and Mapping). However, in traditional SLAM algorithms, the robot is teleoperated by the user, so that it collects as much information as possible for the map. What we are going after for a fully autonomous system is called Active SLAM, where the robot is the one who decides where to go to collect more data to build the map.
Finally, we have the autonomy we desire. What is missing? We need to turn that knowledge into semantic representations so that it can be easily understood by humans. For that, we will enforce computer vision techniques.
Why is it important for Human-Robot-Interaction (HRI)? To put it simple; a system that is able to interact with humans is only limited to its sensor’s range. Being able to move around, not only it’s more practical in terms of interactions and functionalities, but it’s also more natural in order to mimic human-like behavior.
How can cloud help? As many of you might know, cloud offers various advantages over traditional on-board systems. One important aspect is the building cost of the robot that is reduced significantly. Cloud gives the robot enhanced capabilities without the need of extra hardware and also enables the cloud intelligence where various robots will be able to exchange information and collaborate.
Sounds good? Put it all together, you have a novel approach for the core of HRI: intelligence in navigation!