10 Benchmarks That Defined Reinforcement Learning
Benchmarks are the shared language of RL research — the scores every new method gets measured against. These ten, spanning three decades, trace the field's arc from toy problems to superhuman play. (See them all in our benchmarks board.)
The hello world
The canonical solve threshold: average 200+ over 100 episodes. If your algorithm can't do this, nothing else matters. It proved that simple control was learnable — and gave the field its first shared yardstick.
Reward shaping works
A step up in complexity with shaped rewards for leg contact and fuel use. LunarLander proved that well-designed dense rewards could tame problems sparse rewards couldn't touch.
Pixels become playable
DeepMind's DQN learned directly from pixels on 49 of 57 Atari games. It proved end-to-end deep RL was possible — no hand-crafted features — and launched the modern era.
The improvements compose
Rainbow combined six DQN extensions and more than doubled median performance. It proved the Atari benchmark still had headroom, and that algorithmic ingredients stack.
The benchmark falls
Agent57 exceeded the human baseline on all 57 Atari games — including the brutal exploration games like Montezuma's Revenge. It proved that with enough curiosity-driven exploration, the entire suite was solvable.
Intuition becomes computable
AlphaGo's victory proved that Monte Carlo tree search plus deep networks could master a game with more positions than atoms in the universe — a result most experts had placed decades away.
Tabula rasa wins
Starting from random play and learning purely by self-play, AlphaZero crushed the strongest chess engine in history. It proved domain knowledge was optional — search plus learning was enough.
Real-time strategy falls
Grandmaster level in StarCraft II — imperfect information, long horizons, real-time control. It proved deep RL could handle the messiness of real-time decision-making, not just turn-based board games.
Teamwork emerges
Five neural networks beat the Dota 2 world champions, demonstrating long-horizon cooperation: the agents learned strategies spanning tens of thousands of timesteps. It proved multi-agent RL could produce genuine teamwork.
Sim-to-real crosses over
OpenAI's Dactyl solved a Rubik's cube with a dexterous robot hand trained entirely in simulation. It proved domain randomization could bridge the reality gap — policies learned in sim could survive contact with the physical world.
Benchmarks don't just measure progress; they direct it. Each of these told the field what to try next. The question now is which environment will define the next decade — our money is on the GPU-native suites. Browse the directory and place your own bets.