High-quality virtual environments are essential for training and benchmarking reinforcement learning agents that can perform high level skills (executed as a sequence of low-level actions).  Caglar Gulcehre and Tom Le Paine from DeepMind research created and open-sourced the Hard Eight task suite, a collection of challenging exploration problems in partially observable environments with highly variable initial conditions and sparse rewards. The tasks (set in a procedurally generated 3D world) require an agent to use objects (e.g. a wall sensor, blocks) in its environment to access a large apple, which provides a reward. The Hard-Eight Task Suite is designed to make it nearly impossible for an agent to just memorize an open loop sequence of actions to achieve its goal.