Atari Kangaroo Environment

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

Performances of RL Agents

We list various reinforcement learning algorithms that were tested in this environment. These results are from RL Database. If this page was helpful, please consider giving a star!

Star

Human Starts

Result Algorithm Source
12185.0 Prioritized DDQN (rank, tuned) Prioritized Experience Replay
11204.0 DDQN (tuned) Deep Reinforcement Learning with Double Q-learning
10841.0 Rainbow Rainbow: Combining Improvements in Deep Reinforcement Learning
10334.0 DuDQN Dueling Network Architectures for Deep Reinforcement Learning
10241.0 Prioritized DDQN (prop, tuned) Prioritized Experience Replay
9555.5 Distributional DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
9053.0 Prioritized DQN (rank) Prioritized Experience Replay
6138.0 DDQN Deep Reinforcement Learning with Double Q-learning
2739.0 Human Massively Parallel Methods for Deep Reinforcement Learning
2696.0 DQN Massively Parallel Methods for Deep Reinforcement Learning
1431.0 Gorila DQN Massively Parallel Methods for Deep Reinforcement Learning
861.0 PDD DQN Dueling Network Architectures for Deep Reinforcement Learning
125.0 A3C LSTM Asynchronous Methods for Deep Reinforcement Learning
106.0 A3C FF 1 day Asynchronous Methods for Deep Reinforcement Learning
100.0 Random Massively Parallel Methods for Deep Reinforcement Learning
94.0 A3C FF Asynchronous Methods for Deep Reinforcement Learning

No-op Starts

Result Algorithm Source
15487 IQN Implicit Quantile Networks for Distributional Reinforcement Learning
15356 QR-DQN-1 Distributional Reinforcement Learning with Quantile Regression
15227 NoisyNet DuDQN Noisy Networks for Exploration
14854.0 DuDQN Dueling Network Architectures for Deep Reinforcement Learning
14847 DuDQN Noisy Networks for Exploration
14780 QR-DQN-0 Distributional Reinforcement Learning with Quantile Regression
14637.5 Rainbow Rainbow: Combining Improvements in Deep Reinforcement Learning
13651.0 DDQN Deep Reinforcement Learning with Double Q-learning
13349.0 Reactor ND The Reactor: A fast and sample-efficient Actor-Critic agent for Reinforcement Learning
12992.0 DDQN A Distributional Perspective on Reinforcement Learning
12909.0 Distributional DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
12853 C51 A Distributional Perspective on Reinforcement Learning
11266.5 Reactor The Reactor: A fast and sample-efficient Actor-Critic agent for Reinforcement Learning
10944 NoisyNet DQN Noisy Networks for Exploration
10484.5 Reactor The Reactor: A fast and sample-efficient Actor-Critic agent for Reinforcement Learning
8240.5 IMPALA (deep, multitask) IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures
8166 DQN Noisy Networks for Exploration
7259.0 DQN A Distributional Perspective on Reinforcement Learning
6740 DQN Human-level control through deep reinforcement learning
3150.0 ACKTR Scalable trust-region method for deep reinforcement learning using Kronecker-factored approximation
3035.0 Human Dueling Network Architectures for Deep Reinforcement Learning
3035.0 Human Human-level control through deep reinforcement learning
2549.16 Gorila DQN Massively Parallel Methods for Deep Reinforcement Learning
1792.0 PDD DQN Dueling Network Architectures for Deep Reinforcement Learning
1632.0 IMPALA (deep) IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures
1622 Linear Human-level control through deep reinforcement learning
1604 NoisyNet A3C Noisy Networks for Exploration
1166 A3C Noisy Networks for Exploration
52.0 Random Human-level control through deep reinforcement learning
47.0 IMPALA (shallow) IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures
8.8 Contingency Human-level control through deep reinforcement learning

Normal Starts

Result Algorithm Source
9928.7 PPO Proximal Policy Optimization Algorithm
50.0 ACER Proximal Policy Optimization Algorithm
45.3 A2C Proximal Policy Optimization Algorithm