New research in Machine Learning has taught machines how to identify a door, a door handle, how to open that door, and how to confirm that the task was done properly. All by watching humans doing the same task!

As discussed in previous posts, the machine can already recognize the objects it finds in photos and videos. So researchers are expanding on that technology to create robots that learn by watching humans in an unsupervised fashion.

The Building Blocks in Google Research Labs

Below is an example of a human physically instructing a machine to open a door. This technique appeared to work pretty well, especially when supplemented with machine learning.

Interestingly enough, researchers found that they could allow multiple robots to learn at the same time, this way they can all benefit from the experience and learning from one another. The more machines participating, the faster they all learned.

The Problem

Notice the wire connected to the back of the door in the video above. This line runs to a sensor that tells the robot if the door was actually opened or not. The problems presented here show that machines need a way to learn to do these things in an unsupervised fashion. No physical training by humans should be needed other than visual cues and of course, in real life there are no sensors on the backs of doors.

Machine Learning By Watching

In the demonstration below, we can see how the robot is learning to imitate human motions even though the machine does not have identical body parts. For example, the human has two legs that must bend in order to squat, while the robot has a different setup. Regardless of the physical differences in their bodies, the machine has learned how to best mimic the human with the hardware it has, by watching the human move.

Next, researchers applied the same technology to doors. This allows the machines to learn the actions needed to open a door by watching humans rather than being manually instructed.

Rewards

This technology also allows the machine to determine for itself, if it has completed the task appropriately. If the machine determines that the door was opened (without the use of a sensor), then it uses this experience to create a reward or a goal to help re-enforce more complicated but similar tasks.

The researchers started slow, and then gradually increased the complexity of the scenarios noticing that machines were able to learn better when they learned in small chunks and were allowed to use experience from the past to build on.

In the video below, we can see a researcher changing the orientation of a door and the machine is able to conduct the task every time without any sensors.

Modern science is very close to creating machines that can self-teach and self-reward entirely autonomously.

The future for robots looks bright!

References:
https://research.googleblog.com/2016/10/how-robots-can-acquire-new-skills-from.html

https://research.googleblog.com/2017/07/teaching-robots-to-understand-semantic.html