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.