Search Unity

In February, we launched the first round of the Obstacle Tower Challenge. Since the closing of the round, we have received 2000+ entries from 350+ teams. We want to thank all the participants of the first round, congratulate the top eligible teams moving on to round 2, and share the new round 2 version of Obstacle Tower with the public.

Today, we are starting the second round of the Obstacle Tower Challenge for eligible teams. For a team to make it into this round, they needed to train and submit an agent that could achieve an average score of five on unseen versions of the tower. As we described here, this was no trivial feat. We want to congratulate the top teams that made it this far and also thank our partners at Google Cloud Platform for providing GCP credits to the eligible teams and AICrowd for hosting the challenge.

To go along with this new round of the challenge, we are also releasing version 2.0 of the Obstacle Tower Environment. We’ve expanded the number of floors in the tower from 25 to 100, with these higher floors including many new visual styles, obstacles, puzzles to solve and enemies to avoid. We have also significantly expanded the customizability of the environment for researchers looking to study generalization in AI. Although only the top eligible teams will move on to round 2 of the challenge, we encourage everyone to download and try out the latest version.  After round 2 of the challenge, we plan to open source the Obstacle Tower Environment.

What’s new in the Obstacle Tower Environment v2.0

Expanded available floors

We’ve expanded the Obstacle Tower Environment to further push the agent’s need to generalize across new and unseen combinations and floors. Whereas the original version of the environment had only 25 floors in the tower, this new version has 100. These higher floors contain three new visual styles – Industrial, Modern, and Future.

Not only do the higher floors contain different visual appearances, but they also provide additional challenges. All of the mechanics present in the early floors are still present and expanded upon in difficulty in the higher floors. In addition to those mechanics, the higher floors also contain: enemies to dodge, distracting TVs to avoid, more complex floor layouts with circling paths, and larger rooms on each floor with additional platforming challenges.

New Visual Styles – Industrial, Modern, Future

New Obstacles and Distractions

Customizing the environments through reset parameters

In the original release of Obstacle Tower, it was only possible to change the starting floor and initialization seed from the python api. In the 2.0 release, we have significantly expanded on the number of available parameters which can be customized when resetting the environment. These include the ability to change things like the lighting, visual theme, floor layouts, and room contents on the floors in the tower.

Changes and improvements based on feedback from Round 1

Perhaps most importantly, we were able to make changes based directly on the feedback of the users during Round 1 of the challenge. Many of these consisted of bugs or feature requests which were made by the participants themselves, but some of the changes were based on bugs that were only found by agents learning to maximize their scores in the game. One feedback we received was that the placement of the reset button in puzzle rooms was unintuitive. As such, we have now separated out the block, goal, and reset button positions in these rooms, to make it less likely that the agent will press the reset button by accident.

Available now

The Obstacle Tower Environment natively supports the Unity ML-Agents Toolkit and is available to download here. For general issues or discussion of the environment itself, see our GitHub repo. To learn more about the environment, read our research paper. For those interested in an open source release, we are happy to share that we will be releasing the full source code for the Unity project at the end of Round 2. Our aim is to provide a foundation for researchers and the community to extend the Obstacle Tower environment in order to further advance RL research. In the meantime, we hope that the expanded reset parameters will give researchers a lot to flexibility. We can’t wait to see how you use the environment in new and unexpected ways.

And lastly, we would like to thank all the contributions and testers for helping us improve the Obstacle Tower Environment v2.0.

Round 2 finalists

Congratulations to the teams who are moving on to Round 2! When we first launched the challenge a few months ago, we didn’t know how far to expect participants to be able to make it into the tower. We have been pleasantly surprised to find that the top submissions have been of agents not only able to master the mechanics of finding and utilizing keys, but also the mechanics of pushing blocks to solve puzzles.

Below is the final list. Please note, this may differ from the AICrowd leaderboard due to disqualifications and eligibility for the contest.

Participant Round 1 Average Floors Round 1 Average Reward
unixpickle 16.401 29.881
joe_booth 10.001 16.461
dougm 9.601 15.921
karolisram 8.401 13.321
sova876 8.201 13.121
giadefa 8.001 12.821
wywarren 8.001 12.581
PerInDisguise 7.001 10.681
tatsuyaogawa 6.601 9.721
STAR.Lab 6.001 8.721
tky 6.001 8.661
sungbinchoi 5.601 8.101
ipv6 5.601 8.061
kyunghyunlee 5.601 7.881
denamganai_kevin 5.601 7.861
adamloch 5.401 7.621
rudy_gilman 5.401 7.601
wenyuyangpku 5.401 7.561
oleksandra_fedorova 5.401 7.521
hanschoi86 5.401 7.501
TruthMaker 5.401 7.501
petr 5.401 7.501
BIgG 5.401 7.501
ub 5.401 7.481
Petero 5.401 7.481
duc_nguyen 5.401 7.441
gardenermike 5.201 7.461
kenshi_abe 5.201 7.241
hyochini 5.201 7.221
gr1d 5.201 7.201
Leckofunny 5.201 7.161
steven 5.201 7.141
cit 5.201 7.101
felixlaumon 5.001 6.881
xihe 5.001 6.861
kyushik_min 5.001 6.861
Miffyli 5.001 6.781
HappySlice 5.001 6.781
thesoenke 5.001 6.781
paullewislobo 5.001 6.761
andwetry 5.001 6.761
alex_gomez 5.001 6.761
Parilo 5.001 6.721
banjtheman 5.001 6.721

If you have any questions about the challenge please email us at OTC@unity3d.com. If you’d like to work on this exciting intersection of Machine Learning and Games, we are hiring for several positions, please apply!

Ya no se aceptan más comentarios.

  1. Don’t you think it would be appropriate to have a left and a right eye view from the player perspective as available observations? After all, humans playing these games have already gained their knowledge about 3d perspective etc. from real life before playing 3d games. How many one-eyed creatures have survived evolution? 1d observation image may be ok for Mario type games, but is it really enough for this?

    1. Without knowing much about the Unity ML-Agents Toolkit, I’d say that I could see that being useful. Like maybe ray tracing from both eyes to help establish depth for the AI. I’m just trying to find videos of the attempts. Can’t find any… :(