Overview

There are four different levels. Each of them consist of the mother kangaroo on the bottom floor trying to reach the top floor where her joey is being held captive by some monkeys. On each of the levels, there are monkeys who are throwing apples at the mother kangaroo. Sometimes the apples are thrown so that she must jump over them and sometimes they are thrown so that she must duck. If she gets face to face with one of the monkeys, she can punch the monkey with a boxing glove. She can also punch and destroy apples if they’re thrown in level with her gloves. Also, there are pieces of fruit that she can jump up and get for points. Additionally, there is at least one bell on each level that she can hit so that more fruits will appear. She must be wary of the big Ape, who will occasionally appear and try to take her gloves away from her. The level must be completed before the time runs out, otherwise the player will lose a life.

Levels 1, 2 and 4 consist of different platforms that the mother kangaroo must jump onto or climb onto via a ladder. On the third level, the cage in which the kid kangaroo is imprisoned is held up by an entire troop of monkeys and there is a horde of apples that the monkey will unleash if five of them climb up there. On this level, the mother kangaroo must punch each monkey in the stack several times until the cage is lowered and when the cage has been lowered enough, the mother kangaroo must climb to the next floor to get to the kid kangaroo before the cage is raised again or before the monkeys have an avalanche of apple cores unleashed.

Kangaroo has a number of clearly visible glitches in the graphics, such as sprites briefly flickering.[

Description from Wikipedia

State of the Art

Human Starts

Result Method Type Score from
12185.0 PERDDQN (rank) DQN Prioritized Experience Replay
11204.0 DDQN DQN Deep Reinforcement Learning with Double Q-learning
10841.0 RainbowDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
10334.0 DuelingDDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
10241.0 PERDDQN (prop) DQN Prioritized Experience Replay
9555.5 DistributionalDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
9053.0 PERDQN (rank) DQN Prioritized Experience Replay
7465.0 NoisyNetDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
4496.0 DQN2015 DQN Dueling Network Architectures for Deep Reinforcement Learning
2739.0 Human Human Massively Parallel Methods for Deep Reinforcement Learning
2696.0 DQN2015 DQN Massively Parallel Methods for Deep Reinforcement Learning
1431.0 GorilaDQN DQN Massively Parallel Methods for Deep Reinforcement Learning
861.0 DuelingPERDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
577.5 ApeX DQN DQN Distributed Prioritized Experience Replay
125.0 A3C LSTM PG Asynchronous Methods for Deep Learning
106.0 A3C FF (1 day) PG Asynchronous Methods for Deep Learning
100.0 Random Random Massively Parallel Methods for Deep Reinforcement Learning
94.0 A3C FF (4 days) PG Asynchronous Methods for Deep Learning

No-op Starts

Result Method Type Score from
16200.0 PER DQN Scalable trust-region method for deep reinforcement learning using Kronecker-factored approximation
16200.0 PERDDQN (rank) DQN Dueling Network Architectures for Deep Reinforcement Learning
15227.0 NoisyNet-DuelingDQN DQN Noisy Networks for Exploration
14854.0 DuelingDDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
14847.0 DuelingDQN DQN Noisy Networks for Exploration
14681.5 DQfD Imitation Deep Q-Learning from Demonstrations
14637.5 RainbowDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
14492.0 PERDDQN (prop) DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
13651.0 DDQN DQN Deep Reinforcement Learning with Double Q-learning
13567.3 DuelingPERDDQN DQN Deep Q-Learning from Demonstrations
13150.0 DDQN+PopArt DQN Learning values across many orders of magnitude
12992.0 DDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
12909.0 DistributionalDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
12853.0 C51 Misc A Distributional Perspective on Reinforcement Learning
12117.0 NoisyNetDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
10944.0 NoisyNet-DQN DQN Noisy Networks for Exploration
8166.0 DQN DQN Noisy Networks for Exploration
7259.0 DQN2015 DQN Dueling Network Architectures for Deep Reinforcement Learning
6740 DQN2015 DQN Human-level control through deep reinforcement learning
3150.0 ACKTR PG Scalable trust-region method for deep reinforcement learning using Kronecker-factored approximation
3035 Human Human Human-level control through deep reinforcement learning
2549.16 GorilaDQN DQN Massively Parallel Methods for Deep Reinforcement Learning
1792.0 DuelingPERDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
1622 Linear Misc Human-level control through deep reinforcement learning
1604.0 NoisyNet-A3C PG Noisy Networks for Exploration
1416.0 ApeX DQN DQN Distributed Prioritized Experience Replay
1166.0 A3C PG Noisy Networks for Exploration
52 Random Random Human-level control through deep reinforcement learning
8.8 Contingency Misc Human-level control through deep reinforcement learning

Normal Starts

Result Method Type Score from
9928.7 PPO PG Proximal Policy Optimization Algorithms
50.0 ACER PG Proximal Policy Optimization Algorithms
45.3 A2C PG Proximal Policy Optimization Algorithms