LARRUZFounder | Author & Creator of BYOAI

Digital Creatures Learn to Navigate in 3D | Two Minute Papers #153

14 May 2017 · 2 min

Watch on YouTube

2:26

Researchers trained digital characters to navigate 3D environments using two controllers: one handles balance and joint movement, the other handles path-finding and avoiding obstacles. The system can dribble a ball and transfer some skills between different tasks.

Questions this video answers

30 times per second.

Trained via a four-layer neural network.

Twice per second.

Trained via a deep convolutional neural network.

Navigation in 3D with static and dynamic moving obstacles and tasks like dribbling a ball toward a target.

Chapters

Dear fellow scholars, this is two minute papers with earlier we have talked about a few amazing algorithms to teach digital creatures to walk, and this time we are interested in controlling the joints of a digital character to not only walk properly, but take into consideration its surroundings. This new version can navigate in 3D with static and dynamic moving obstacles or even dribble a ball toward a target. Loving the execution and the production value of this paper, this is accomplished by an efficient system that consists of two controllers that are represented by learning algorithms. 1. The low level controller is about maintaining balance and proper limb control by manipulating the joint positions and velocities appropriately. This controller operates on a fine timescale 30 times per second and is trained via a four layer neural network. 2. The high level controller can accomplish bigger overarching goals such as following a path or avoiding static and dynamic obstacles. We don't need to run this so often, therefore, to save resources. This controller operates on a coarse timescale only twice each second and is trained via a deep convolutional neural network. It also has support for a small degree of transfer learning. Transfer learning means that after successfully learning to solve a problem, we don't have to start from scratch for the next one, but we can reuse some of that valuable knowledge and get a head start. This is a heavily researched area and is likely going to be one of the major next frontiers in machine learning research. Now, make no mistake, it is not like transfer learning is suddenly to be considered a solved problem, but in this particular case, it is finally a possibility. Really cool. I hope this brief expose fired you up too. This paper is a bomb. Make sure to have a look. As always, the link is available in the video description. And by the way, with your support on Patreon, we will soon be able to spend part of our budget on empowering research projects. How amazing is that? The new two Minute Papers shirts are also flying off the shelves. Happy to hear you're enjoying them so much. If you're interested, hit up 2minutepapers.com if you are located in the US, the EU and worldwide stores. Link is available in the video description. Thanks for watching and for your generous support and I'll see you next time.