AI Habitat, serving as a platform for research in the diverse forms of Artificial Intelligence, recently hosted the Habitat Challenge in which the SemExp system overcame Samsung and took first place in the ObjectNav category of the challenge.
The system took the use of automated machine learning, which is a deliberate field of study, and taught how to recognize familiar objects. This is something to look forward to in robotics, as the development of such a trait is bound to have alternating impacts.
In an example video shot by CMU, titled as “Common Sense Guides Robot Around The House”, machine learning is seen at its pinnacle. Looking like a real-life Wall-E, the robot tries to find the couch using his observation and analysis.
We can notice in the video how our robotic little friend is able to point out key differences between most objects distinctively. For instance, it knows the difference between a kitchen table and an end table. In this regard, the robot is keen on finding his objective in the best possible room he deduces he’s going to be successful in.
CMU says that the system of machine learning could make for an easier time with robots in the not so distant future. They go on to explain how it would be more practical in commanding the robot to “get the remote control next to the plant”.
We can further perceive in the example video of how the robot formulates a predictive semantic map, allowing him to make choices, and act accordingly.
Ph.D. student of Machine Learning Devendra S. Chaplot said that it’s nothing but common sense, that if you were to go looking for a fridge, you wouldn’t go to the bathroom, but straight to the kitchen where the probabilities are much much higher.
“Classical robotic navigation systems, by contrast, explore a space by building a map showing obstacles. The robot eventually gets to where it needs to go, but the route can be circuitous.”
CMU has affirmed that this was not by far the first venture towards teaching semantic navigation to a robot. Instead, previous attempts focused too much on the actual memorization of objects concerning their location. This time, however, the concentration was directed to linking an object with its most expected location, and this ideology is more likely to birth success.