A Survey on Model-Free Goal Recognition

Leonardo Amado, Sveta Paster Shainkopf, Ramon Fraga Pereira, Reuth Mirsky, Felipe MeneguzziIJCAI 2024

Abstract

Goal Recognition (GR) is the task of inferring an agent's intentions by observing a sequence of their actions. Traditionally, this has required hand-coded "domain theories" or expert models, which are often difficult to construct for complex, real-world environments. This survey explores the emerging paradigm of "model-free" goal recognition, categorizing modern techniques that learn to recognize intentions directly from data, including Deep Learning, Reinforcement Learning, and Neuro-Symbolic AI.

1. Introduction

Goal Recognition is a fundamental capability for intelligent agents, enabling them to anticipate actions and collaborate effectively. Historically, GR relied on formal models (e.g., PDDL) to define possible actions and environmental changes. While precise, these model-based approaches are brittle when faced with noisy sensor data or unexpected behaviors. The "Model-Free" paradigm addresses these limitations by enabling agents to recognize goals without a predefined, explicit model, instead utilizing observed traces, raw images, or interactive simulations.

2. Taxonomy of Model-Free Goal Recognition

The landscape of model-free GR can be organized across several key dimensions, as synthesized in the comprehensive comparison of literature provided in Table 1.

2.1 Model Affinity

Approaches are primarily distinguished by their reliance on internal representations:

  • Agnostic: These systems map observations directly to goals (e.g., using a neural network as a classifier) without building an internal model of the environment's rules.
  • Model-Approximate: These systems first learn an approximate model of the domain from data and subsequently use that model to reason about goals.

2.2 Core Techniques

The survey identifies four main pillars of model-free GR:

  • Supervised Learning: Treating GR as a multi-class classification problem using architectures like LSTMs or CNNs.
  • Reinforcement Learning (RL): Utilizing learned Q-values or policies to represent and evaluate goal hypotheses.
  • Symbolic Methods: Employing techniques such as Process Mining or Combinatory Categorial Grammars (CCG) to identify structural patterns in behavior.
  • Neuro-Symbolic AI: Integrating neural networks for raw data processing with symbolic reasoning for logical inference.
Table 1: A comprehensive comparison of model-free goal recognition papers, categorized by environment, technique, input data, and recognition characteristics.

3. Input Modalities and Data Sources

Modern GR has moved beyond symbolic action logs to incorporate diverse data types:

  • Structured Data: Plan traces consisting of state and action sequences, often derived from classical planning benchmarks.
  • Unstructured Data: Raw pixel data from image sequences or video streams, and natural language descriptions.
  • Simulated Data: Large-scale datasets sampled from Reinforcement Learning environments or game engines.

4. Key Insights and Discussion

The transition to model-free approaches highlights several critical trade-offs and trends:

  • The Agnostic vs. Approximate Trade-off: Agnostic models offer high speed and handle raw sensor data well but require massive datasets and lack interpretability. Model-approximate approaches are more data-efficient and provide "explanations" but are constrained by the quality of the learned model.
  • The Neuro-Symbolic Frontier: Promising recent works (e.g., PPR, GRAQL) attempt to bridge the gap by using neural networks for noisy inputs while maintaining a symbolic backbone for logical consistency.
  • From Offline to Online Recognition: There is a significant shift toward online recognition—inferring goals incrementally as actions unfold—which is essential for real-time human-robot interaction.
  • Abductive Reasoning: While often treated as classification (deduction), GR is fundamentally a problem of abduction—finding the best explanation for a set of observations.

5. Challenges and Future Work

Despite significant progress, the field faces several ongoing challenges:

  • Data Requirements: Deep Learning models remain highly dependent on large labeled datasets, which are often unavailable in specialized domains.
  • Explainability: The "black box" nature of many model-free approaches is a barrier for safety-critical applications.
  • Generalization: Most current models are domain-specific; transferring a recognizer between different environments remains difficult.
  • The Role of LLMs: Large Language Models (LLMs) represent a potential new class of model-approximate recognizers that can leverage natural language to understand world dynamics, potentially reducing the need for expert-engineered models.

6. Conclusion

Model-free goal recognition represents a vital shift toward more flexible and scalable AI. By leveraging data-driven techniques, the field is overcoming the limitations of expert-coded models, paving the way for intelligent systems that can operate effectively in complex, unmodeled environments.

Summary updated

BibTeX
@inproceedings{amado2024survey,
  author = {Leonardo Amado and Sveta Paster Shainkopf and Ramon Fraga Pereira and Reuth Mirsky and Felipe Meneguzzi},
  title = {{A Survey on Model-Free Goal Recognition}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure