Your submission is now a Draft.

Once it's ready, please submit your draft for review by our team of Community Moderators. Thank you!

You have been invited to co-author this question.

When it is ready, the author will submit it for review by Community Moderators. Thanks for helping!


This question now needs to be approved by community moderators.

You have been invited to co-author this question.

It now needs to be approved by community moderators. Thanks for helping!

When will an AI system match amateur human performance on Obstacle Tower with no domain-specific hardcoded knowledge?


Following the success of agents on classical AI benchmarks like chess, game engine company Unity has made a new video game called 'Obstacle Tower', specifically designed to showcase weaknesses of current Deep RL algorithms.

The point of Obstacle Tower is to make it as high as possible up a tower. In order to do so, the agent must…

  • Learn directly from pixel input to solve...

  • ...low level control problems...

  • ...and high-level planning problems...

  • ...using a sparse reward signal

In addition, Obstacle Tower is designed to contain:

  • Physics-driven interactions. The movement of the agent and other objects within the environment are controlled by a real-time 3D physics system.

  • High visual fidelity. The graphics are much closer to photo-realistic than other platforms such as the Atari Learning Environment, VizDoom, or DeepMind Lab.

  • Procedural generation of nontrivial levels and puzzles. Instead of a stock collection of levels, new ones can be created automatically.

  • Procedurally varied graphics. The same level might look different on different runs due to changes in textures, lighting conditions and object geometry.

So to better understand the future success of deep reinforcement learning (and potentially other techniques), we ask:

When will an AI system achieve a mean performance on Obstacle Tower of at least 9 floors solved in a single episode, in the Strong Generalization condition, using no domain-specific hardcoded knowledge?


See tables 1 and 2 in the original Unity paper.

Human performance (from volunteer Unity Technologies employees) had mean scores (and standard deviations) 12 (6.8) floors reached in the training condition and 9.3 (3.1) in the test condition, and maximum of 20 in the test condition.

The best scores using state-of-the-art algorithms were 0.6 (0.8) with OpenAI’s Proximal Policy Optimisation and 1.6 (0.5) with DeepMind’s Rainbow DQN.

Make a Prediction


Note: this question resolved before its original close time. All of your predictions came after the resolution, so you did not gain (or lose) any points for it.

Note: this question resolved before its original close time. You earned points up until the question resolution, but not afterwards.

Current points depend on your prediction, the community's prediction, and the result. Your total earned points are averaged over the lifetime of the question, so predict early to get as many points as possible! See the FAQ.