Interpretable Latent Spaces for Learning from Demonstration

Yordan Hristov, Alex Lascarides, Subramanian RamamoorthyCoRL 2018

Abstract

Bridging the gap between high-level human language and low-level robotic perception remains a significant challenge in robotics. This paper introduces a framework that enforces interpretability within a robot's latent representation by aligning specific dimensions with human-centric concepts like color and shape. Utilizing a modified $\beta$-VAE architecture with auxiliary classification losses, the system "untangles" these features, enabling the robot to ground instructions more effectively and recognize novel, "unknown" objects in an open-world setting.

Introduction

When a human instructs a robot to "pick up the red cube," the robot must translate abstract symbols into a high-dimensional stream of raw pixels. While standard deep learning approaches like Variational Autoencoders (VAEs) excel at data compression, their internal "latent spaces" are often uninterpretable black boxes. A single dimension might represent an inseparable mix of visual features, making it difficult to debug errors or trust the robot's decision-making process. Furthermore, traditional models often fail to identify "unknown unknowns," confidently mislabeling novel objects instead of acknowledging uncertainty.

Methodology

The proposed framework builds upon the $\beta$-VAE architecture, introducing Auxiliary Classification Losses to ensure latent space interpretability.

Axes Alignment and the Single Dimension Constraint

The core innovation lies in forcing a one-to-one mapping between human concepts (e.g., color, shape) and specific latent dimensions. To achieve this, the authors append linear classifiers to the network, each restricted to observing only one specific dimension of the latent space. For example, if the "color" classifier is limited to dimension $z_0$, the network must consolidate all color-related information into that single axis to minimize classification loss.

The proposed framework uses a $\beta$-VAE architecture with auxiliary classification losses to align latent dimensions with human-interpretable concepts like color and shape.

Weak Supervision and Open-World Recognition

The model utilizes a weighted loss function comprising reconstruction loss ($\alpha$), KL divergence ($\beta$), and classification loss ($\gamma$). This allows for weak supervision, where not every training image requires a label. By modeling the distribution of known objects as 1D Normal distributions along the aligned axes, the robot can identify "unknown" objects. If a latent value falls into the "no-man's-land" between known clusters, the system flags the object as unknown rather than forcing it into a pre-defined category.

Experimental Setup

The framework was evaluated on two distinct datasets:

  1. Modified dSprites: A synthetic dataset of 3,500 images featuring simple shapes (squares, hearts, ellipses) in 8 colors.
  2. Real-World Table-top Objects: A dataset of 7,500 image crops from human demonstrations, including juggling balls and "orbs" (Lego blocks). Crucially, certain items—such as a green ball and a yellow duck—were held out during training to test the robot's ability to handle unobserved categories.
The dataset includes labeled and unlabeled objects, allowing the model to identify 'unknown' items like the yellow duck by detecting values in the latent space that fall outside known distributions.

Results and Discussion

The results demonstrate that forcing concepts into independent, orthogonal axes ensures that changes in one attribute (e.g., color) do not interfere with the perception of another (e.g., shape).

  • Alignment Quality: Principal Component Analysis (PCA) revealed that the proposed model achieved significantly lower entropy in its axes alignment compared to standard $\beta$-VAEs, indicating "purer" concept representations.
  • Open-World Robustness: In real-world tests, the model achieved an F1 score of 0.95 for identifying unknown colors, vastly outperforming a standard classifier-only baseline which scored only 0.1.
  • Generalization: The system successfully identified that a "yellow duck" possessed a known color (yellow) but an unknown shape, proving the effectiveness of the disentangled latent space.

Conclusion and Future Work

This research provides a mathematical basis for robots to "admit ignorance" when encountering new objects. While the current approach requires manual mapping of dimensions to concepts and focuses on discrete attributes, it represents a significant step toward transparent and reliable robotic learning. Future work will explore automating dimension discovery and scaling the approach to more complex, continuous concepts in cluttered environments.

Summary updated

BibTeX
@inproceedings{hristov2018interpretable,
  author = {Yordan Hristov and Alex Lascarides and Subramanian Ramamoorthy},
  title = {{Interpretable Latent Spaces for Learning from Demonstration}},
  booktitle = {Conference on Robot Learning},
  year = {2018},
}
Figure