Mind the GAP! The Challenges of Scale in Pixel-based Deep Reinforcement Learning

Scaling deep reinforcement learning in pixel-based environments presents a significant challenge, often resulting in diminished performance. While recent works have proposed algorithmic and architectural approaches to address this, the underlying cause of the performance drop remains unclear.