Feature-Based vs. GAN-Based Learning from Demonstrations: When and Why

Chenhao Li, Marco Hutter, Andreas KrausearXiv 2025 · DOI · arXiv

Introduction

This survey from researchers at ETH Zurich provides a comprehensive "field guide" to the two dominant paradigms in physics-based imitation learning: Feature-Based and GAN-Based methods. As robotics moves from simple repetitive tasks to complex, lifelike movements, understanding the trade-offs between these two approaches is essential for building autonomous agents that learn effectively from expert demonstrations.

The Challenge of Reward Engineering

In Reinforcement Learning (RL), designing a reward function from scratch is notoriously difficult. To solve this, researchers utilize Learning from Demonstrations (LfD), where an agent learns by observing expert data, such as motion capture or teleoperation. Historically, the field has been split into two distinct camps: feature-based methods that explicitly track reference motions, and GAN-based methods that learn an implicit reward through adversarial competition.

Methodological Evolution

The authors categorize the evolution of these imitation learning methods into three distinct stages:

1. Classic Feature-Based (Explicit Rewards)

These methods, exemplified by DeepMimic, rely on "phase variables" to synchronize the agent with a specific reference clip. They utilize dense, per-frame rewards based on hand-crafted features like joint positions and velocities. This approach ensures high-fidelity tracking but often struggles with scalability and generalization to unstructured data.

Architecture of classic feature-based imitation learning using explicit state tracking.

2. Classic GAN-Based (Implicit Rewards)

Methods like Adversarial Motion Priors (AMP) treat the policy as a generator in a minimax game. A discriminator network analyzes short snippets of motion to judge realism, providing an implicit reward. While this removes the need for time-synchronization, it introduces significant training instability, such as mode collapse.

3. The Convergent Paradigm (Structured Representations)

The latest wave of research sees both camps converging toward the use of latent embeddings. By mapping complex motions into a structured "motion space" (often via VAEs or self-supervised learning), feature-based methods can handle diverse datasets, while GAN-based methods can avoid the saturation and collapse issues that plagued earlier versions.

Key Insights

  • The Myth of "No Hand-Crafting": A common belief is that GANs eliminate the need for hand-crafted features. However, the authors note that discriminators still require carefully selected state features to prevent them from becoming too effective too quickly, which would leave the agent without useful feedback.
  • Dataset Bias vs. Algorithmic Merit: "Naturalness" in motion is often a property of the dataset rather than the algorithm itself. An algorithm's true merit lies in its training stability, scalability, and ability to handle unstructured data.
  • The Hybrid Future: The most effective modern systems are increasingly hybrid, combining the interpretability of feature-based rewards with the scalability of learned latent spaces.

Comparative Analysis

The choice between these paradigms is no longer binary, but depends on the specific requirements of the task. The following table summarizes the fundamental trade-offs identified in the survey:

Systematic comparison of algorithmic trade-offs between GAN-based and feature-based paradigms.

Conclusion and Future Directions

While structured representations have bridged the gap between these two paradigms, challenges remain. Feature-based methods still struggle to deviate from references when tasks require environmental adaptation (e.g., stepping over an obstacle). Furthermore, implementing modern latent-conditioned systems is significantly more complex than their "vanilla" predecessors. The next frontier involves integrating high-level planning, such as diffusion models, with these low-level imitation controllers to create robots that are both lifelike and functional in real-world scenarios.

Summary updated

Abstract

This survey provides a comparative analysis of feature-based and GAN-based approaches to learning from demonstrations, with a focus on the structure of reward functions and their implications for policy learning. Feature-based methods offer dense, interpretable rewards that excel at high-fidelity motion imitation, yet often require sophisticated representations of references and struggle with generalization in unstructured settings. GAN-based methods, in contrast, use implicit, distributional supervision that enables scalability and adaptation flexibility, but are prone to training instability and coarse reward signals. Recent advancements in both paradigms converge on the importance of structured motion representations, which enable smoother transitions, controllable synthesis, and improved task integration. We argue that the dichotomy between feature-based and GAN-based methods is increasingly nuanced: rather than one paradigm dominating the other, the choice should be guided by task-specific priorities such as fidelity, diversity, interpretability, and adaptability. This work outlines the algorithmic trade-offs and design considerations that underlie method selection, offering a framework for principled decision-making in learning from demonstrations.
BibTeX
@misc{li2025feature,
  author = {Chenhao Li and Marco Hutter and Andreas Krause},
  title = {{Feature-Based vs. GAN-Based Learning from Demonstrations: When and Why}},
  year = {2025},
  eprint = {2507.05906},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.05906},
}
Figure