AIBIRDS: The Angry Birds Artificial Intelligence Competition

Jochen RenzAAAI 2015 · DOI

Abstract

The AIBIRDS competition challenges researchers to develop AI agents capable of mastering Angry Birds, a physics-based simulation game. Unlike traditional games with discrete states and perfect information, Angry Birds requires reasoning about continuous action spaces and uncertain physical outcomes. This paper outlines the competition's framework, the technical challenges of physics-based simulation games, and the progress of AI agents in matching human performance.

Introduction

While AI has conquered classic board games like Chess and Go, these environments are "perfect information" games with discrete moves and predictable outcomes. The next frontier for AI is Physics-Based Simulation Games (PBSGs). In these environments, agents must interact with a world that is messy and unpredictable, much like the real world.

In the context of Angry Birds, the challenge is three-fold:

  1. Continuous Action Space: A player must choose a precise release point $(x, y)$ and timing $(t)$ for special abilities. In theory, the number of possible actions is infinite.
  2. The Simulation Gap: To predict the outcome of a shot, an agent needs to know internal parameters like mass, friction, and gravity. However, agents (like humans) only see the screen. Without these parameters, perfect simulation is impossible.
  3. Qualitative Reasoning: Humans are naturally gifted at "physical intuition"—knowing that hitting a specific support beam will cause a structure to collapse. Replicating this "if I do this, then that happens" logic in an AI remains a massive hurdle in general intelligence.
A typical Angry Birds level showcasing the complex physical environment agents must navigate, including various materials and structural vulnerabilities.

Methods

The competition provides a framework where AI agents interact with the game via a Chrome browser extension. To level the playing field, agents only receive screenshots of the game, ensuring they have the same information as a human player.

Technical Components

The organizers provide several baseline modules to lower the barrier to entry:

  • Computer Vision (CV) Module: Detects object boundaries and identifies materials within the scene.
  • Trajectory Planning Module: Maps a desired target point to the necessary slingshot release coordinates.

Participant Strategies

Successful teams have moved beyond simple "Naive Agents" to use more sophisticated approaches:

  • Logic Programming and Qualitative Reasoning: Modeling the structural stability of the pig's fortresses to identify weak points.
  • Advanced Simulation: Running internal "best guess" physics models to predict damage before taking a shot.
  • Machine Learning: Using Bayesian ensemble learning to select strategies based on the specific layout of a level.

Key Results and Insights

As of the paper's publication, 36 teams from 16 countries had participated. The evaluation metrics rely on benchmarking against both standard and unknown levels, as well as direct "Human vs. Machine" challenges.

  • Closing the Gap: AI performance is improving rapidly. In 2013, agents could beat human beginners. By 2014, the best AI agents were performing in the top third of human players.
  • Strategy Selection: A core insight is that the outcome of an action in PBSGs is only known after execution. This forces a shift from exhaustive search trees to predictive modeling. Interestingly, winning agents often use a library of different strategies and a meta-logic layer to choose the best one for a given level layout.
  • Human-Like Information Processing: By restricting agents to visual input, the competition forces researchers to solve computer vision and spatial reasoning problems simultaneously with strategic planning.

Limitations and Future Work

Despite the progress, AI agents still struggle to match the "best" human players, particularly in levels requiring highly creative or multi-step physical "trick shots."

The authors propose several future directions to expand the scope of the research:

  • AIBIRDS Snap!: A version using a visual programming language designed to teach children AI concepts.
  • Game Theory Track: A new competitive mode where two agents "bid" points for the right to take the first shot, adding economic and strategic risk assessment to the physics problem.
  • Generalization: The ultimate goal remains solving the problem of reasoning in unknown physical environments, a skill essential for the development of advanced robotics and autonomous systems.

Summary updated

BibTeX
@inproceedings{renz2015aibirds,
  author = {Jochen Renz},
  title = {{AIBIRDS: The Angry Birds Artificial Intelligence Competition}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2015},
  doi = {10.1609/aaai.v29i1.9347},
}
Figure