The 2017 AIBIRDS Competition

Matthew Stephenson, Jochen Renz, Xiaoyu Ge, Peng ZhangarXiv 2018 · OpenReview

Abstract

This paper presents an overview of the sixth AIBIRDS competition, held at the 26th International Joint Conference on Artificial Intelligence. This competition tasked participants with developing an intelligent agent which can play the physics-based puzzle game Angry Birds. This game uses a sophisticated physics engine that requires agents to reason and predict the outcome of actions with only limited environmental information. Agents entered into this competition were required to solve a wide assortment of previously unseen levels within a set time limit. The physical reasoning and planning required to solve these levels are very similar to those of many real-world problems. This year's competition featured some of the best agents developed so far and even included several new AI techniques such as deep reinforcement learning. Within this paper we describe the framework, rules, submitted agents and results for this competition. We also provide some background information on related work and other video game AI competitions, as well as discussing some potential ideas for future AIBIRDS competitions and agent improvements.

Introduction

The AIBIRDS competition utilizes the popular game Angry Birds as a benchmark for a fundamental challenge in artificial intelligence: physical reasoning in unknown environments. Unlike games like Chess or Go, which operate on discrete states and perfect information, Angry Birds presents a continuous state/action space governed by a sophisticated physics engine.

To succeed, an agent must process raw pixels via computer vision to identify objects, reason about physical properties such as mass, friction, and stability, and predict action outcomes without knowing the exact underlying parameters. This complexity mirrors the challenges faced by real-world robots interacting with physical objects.

Methodology

The 2017 competition featured 10 agents utilizing three primary architectural approaches:

  • Heuristic-Based Agents: These agents, including the winner Eagle’s Wing, rely on pre-defined strategies like "target the pig" or "destroy the base of the structure." Their computational efficiency allows for a high volume of shot attempts within competition time limits.
  • Simulation-Based Agents: Agents such as IHSEV and Angry-HEX utilize internal physics engines (e.g., Box2D) to simulate thousands of potential shots, selecting the one with the highest predicted damage. While precise, these agents are slower and dependent on the accuracy of their internal models.
  • Reinforcement Learning (RL): This year introduced deep RL agents. AngryBNU employed Deep Deterministic Policy Gradients (DDPG) to predict optimal shots based on visual features.

All agents interacted with the game through a standardized server-client architecture, receiving screenshots and returning slingshot coordinates to the game engine.

The server-client architecture used to isolate the AI agents from the game engine.

Evaluation and Datasets

Agents were evaluated across several level categories:

  • Benchmark Sets: Two sets of 21 levels from the "Poached Eggs" episode were provided for training and tuning.
  • Competition Levels: 32 secret levels, either manually designed or procedurally generated, were used for the tournament rounds.
  • Man vs. Machine Levels: Four highly difficult levels designed specifically to test "creative" reasoning and multi-step planning.

Key Results and Discussion

The competition results highlighted both the progress of AI agents and the remaining gap compared to human performance.

Competition scores for the top 10 agents across qualification, quarter-finals, semi-finals, and the grand final.

Eagle’s Wing (University of Alberta/Zazzle Inc.) secured first place, proving the robustness of well-tuned heuristic strategies. IHSEV (France) followed in second, demonstrating the competitiveness of simulation-based approaches when computation time is managed effectively.

Several key insights emerged from the tournament:

  • The "Greedy" Trap: Most agents prioritize immediate damage, lacking the intuition to perform non-obvious setup shots that lead to larger collapses later.
  • Speed as a Factor: Given the 30-minute round limit, faster heuristic agents often outscored slower, more "intelligent" simulation agents by simply taking more shots.
  • RL Overfitting: While deep RL agents performed well on benchmarks, they struggled with new levels, suggesting they overfit to specific visual layouts rather than learning general physical principles.
  • Human Superiority: In the "Man vs. Machine" challenge, the top human score (178,290) was more than double the top AI score (73,540). Humans naturally excel at the multi-step planning required for complex levels.

Conclusion and Future Work

The 2017 AIBIRDS competition underscores the need for improved sample efficiency in RL agents, potentially through the use of massive procedurally generated training sets. Future research should focus on moving beyond "greedy" strategies toward multi-shot planning and integrating qualitative physical reasoning with real-time learning to adapt to new game mechanics.

Summary updated

Abstract

This paper presents an overview of the sixth AIBIRDS competition, held at the 26th International Joint Conference on Artificial Intelligence. This competition tasked participants with developing an intelligent agent which can play the physics-based puzzle game Angry Birds. This game uses a sophisticated physics engine that requires agents to reason and predict the outcome of actions with only limited environmental information. Agents entered into this competition were required to solve a wide assortment of previously unseen levels within a set time limit. The physical reasoning and planning required to solve these levels are very similar to those of many real-world problems. This year's competition featured some of the best agents developed so far and even included several new AI techniques such as deep reinforcement learning. Within this paper we describe the framework, rules, submitted agents and results for this competition. We also provide some background information on related work and other video game AI competitions, as well as discussing some potential ideas for future AIBIRDS competitions and agent improvements.
BibTeX
@misc{stephenson2018aibirds,
  author = {Matthew Stephenson and Jochen Renz and Xiaoyu Ge and Peng Zhang},
  title = {{The 2017 AIBIRDS Competition}},
  year = {2018},
  eprint = {1803.05156},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1803.05156},
}
Figure