Search Unity

Today we are announcing the upcoming launch of the Obstacle Tower Challenge: a first-of-its-kind artificial intelligence challenge designed to test the capabilities of intelligent agents and accelerate the research and development of AI. The Obstacle Tower Challenge will be a new competition aimed at testing the vision, control, planning, and generalization abilities of AI agents — capabilities that have yet to be fully tested together.

We built the Challenge on the Obstacle Tower, Unity’s new procedurally generated environment intended for machine learning researchers. The Obstacle Tower is a game-like environment designed to push the boundaries of the human-machine relationship by examining how machines operate in a variety of areas, including computer vision, locomotion skills, and high-level planning – skills that are easy for a human, but for a machine, take learning, practice, and testing.

It is important for the development and advancement of AI methods to have a good benchmark, so that performance and achievements can be fairly and easily compared. This is why we built the Obstacle Tower and why we’re launching our first challenge. We are hoping a little friendly competition will help stimulate AI research and further the studies and creations in reinforcement learning.

The Obstacle Tower environment natively supports the Unity ML-Agents Toolkit and available to download today. We also invite you to read the research paper for more information.

The Obstacle Tower Challenge officially begins on February 11, 2019, at 00:00:01 PST. At that time, entrants can review all the rules and regulations, download our Starter Kit and begin training their agents. Participants will have the opportunity to win prizes in the form of cash, travel vouchers, and Google Cloud Platform credits, valued at over $100,000.

We’ll check back in two weeks for our kick-off. For now, we encourage everyone to mark their calendars and learn more about the Tower!

13 コメント

コメントの配信登録

返信する

これらの HTML タグや属性を使用できます: <a href=""> <b> <code> <pre>

  1. In the paper they let AI discover and train totally from scratch without any rules. In reality you can define rules like, if you see a green door, then go for it. This simplifies a lot the problem. Who says rule-based algorithms are not AI?

    1. Arthur Juliani

      2月 10, 2019 3:05 am 返信

      Hi test,

      This is a good question. We are particularly interested in learning-based AI systems. This is because while it is easy to define rules in games and end up with an “AI,” those same rule-based systems don’t work well in the real world, which is much messier. Many researchers are interested in learning systems which can be trained in games and simulations and then applied to the physical world. The idea behind Obstacle Tower is to help contribute to this line of research.

  2. Train AI with image input? Is it even possible for now? In that research paper you provide, AI only beat 5 floor , then we are supposed to achieve more?If so, This challenge is really high level.

    1. Arthur Juliani

      2月 10, 2019 3:06 am 返信

      Hi one cup,

      You are right. This is a difficult challenge! We have particularly built it to help push the development of new algorithms and methods. While it is true that the agents we trained are so far only able to make it to an average test floor of 5, we think that the community will be able to make a lot more progress than that in the coming months and years.

  3. Tayler Kirk Allen

    1月 30, 2019 4:13 pm 返信

    Which versions of the ML-Agents Toolkit are supported? I am using Linux and cannot currently use version 0.6.0. Also when will there be more information on how we should submit our agents to the contest? I use the python api in combination with Pytorch and I need to be certain this is supposed.

    Thank you.

    1. Arthur Juliani

      1月 31, 2019 1:44 am 返信

      Hi Tayler. The current version uses ML-Agents v0.6. Can you share why that doesn’t work on linux for you? We’ve been using it on linux quite a lot internally without issues.

      As for the contest, and code you submit that can be put into a docker container and interact with our python API will be allowed. This includes Pytorch. Looking forward to your submission.

  4. Very Welcome and this is the perfect opportunity to re-study Unity.

  5. is everyone allowed to participate or are there any specific countries that people can’t participate from?

    1. We intend to allow people from all over the world to participate. And the team is working hard to make this happen. :)

      1. Thanks! Also everyone thats is participating (regardless of the country) can win any of the prizes ?

  6. Please tell me you plan on twitch streaming this

    1. Great idea! ;)

  7. Lars Steenhoff

    1月 28, 2019 3:10 pm 返信

    https://www.youtube.com/watch?v=hGNxXqQgHYY

    Reminds me of mouse maze challenges