SVM-based Transfer of Visual Knowledge Across Robotic Platforms

Jie Luo, Andrzej Pronobis, Barbara CaputoInfoscience 2019 · DOI

Abstract

This paper presents an SVM-based algorithm for the transfer of knowledge across robot platforms aiming to perform the same task. Our method exploits efficiently the transferred knowledge while updating incrementally the internal representation as new information is available. The algorithm is adaptive and tends to privilege new data when building the Support Vector (SV) solution. This prevents the old knowledge from nesting into the model and eventually becoming a possible source of misleading information. We tested our approach in the domain of vision-based place recognition. Extensive experiments show that using transferred knowledge clearly pays off in terms of performance and stability of the solution.

Introduction

In autonomous robotics, vision-based place recognition is typically treated as a "tabula rasa" learning problem. Each new robot deployed must learn its environment from scratch, a process that is both data-intensive and time-consuming. This paper addresses the challenge of knowledge transfer: enabling a new robot (Robot B) to leverage the prior experience of a veteran robot (Robot A).

Transferring knowledge is difficult because robots are rarely identical. Differences in camera height, tilt, and sensor specifications mean that the same room can look significantly different to different platforms. The proposed method aims to use Robot A's data as a starting point while allowing Robot B to adapt its internal model as it gains its own sensory experience.

Methodology

The framework is built upon Support Vector Machines (SVMs) and incorporates three key innovations:

  1. Incremental Learning: To avoid retraining on the entire history of images, the system employs a fixed-partition incremental SVM. It updates the decision boundary by combining Support Vectors (SVs) from the previous model with new data batches.
  2. Knowledge Transfer Initialization: Robot B initializes its classification model using the SVs provided by Robot A. This provides an immediate, though initially imperfect, understanding of the environment.
  3. Adaptive Support Vector Reduction: To maintain memory efficiency, a reduction algorithm prunes redundant SVs. Crucially, the algorithm is modified to privilege new data. If knowledge from Robot A becomes redundant or conflicts with Robot B’s observations, the older data is discarded, ensuring the robot eventually relies on its own perspective.

Experimental Setup

The study utilized the IDOL2 (Image Database for rObot Localization 2) dataset, which includes 24 image sequences across five functional rooms (offices, kitchen, corridor, and printer area) captured over six months under varying lighting conditions.

The experiments involved two distinct robotic platforms with significant physical disparities:

  • PeopleBot (Robot A): Features a camera mounted approximately 98cm from the ground.
  • PowerBot (Robot B): Features a camera mounted at 36cm, tilted upward.
The PeopleBot (Robot A) used as the source platform for knowledge transfer.

This difference in perspective creates a significant "domain shift," testing the algorithm's ability to adapt transferred knowledge to a new viewpoint.

Results and Discussion

Experimental results demonstrate several key advantages of the transfer learning approach:

  • Immediate Performance Gains: Robot B showed a massive accuracy advantage from the very first training examples compared to learning from scratch.
  • Faster Convergence: The transfer-enabled robot reached high accuracy thresholds (above 90%) significantly earlier in its deployment.
  • Memory Efficiency: The SV reduction algorithm successfully prevented model bloat. Over time, the memory footprint of the transfer model became comparable to a model trained from scratch as redundant "teacher" data was pruned.
  • Stability: Transferred knowledge acted as a stabilizer, providing global context that prevents the robot from "forgetting" known rooms while exploring new ones, even in the face of unbalanced data streams.

Conclusion

The proposed SVM-based transfer method effectively bridges the gap between different robotic platforms for place recognition. By treating transferred knowledge as a temporary scaffold that is eventually replaced by the robot's own experience, the system achieves high performance with minimal training time. Future work will explore the use of unlabeled data through semi-supervised learning and the integration of multi-modal sensor data to further enhance robustness.

Summary updated

Abstract

This paper presents an SVM--based algorithm for the transfer of knowledge across robot platforms aiming to perform the same task. Our method exploits efficiently the transferred knowledge while updating incrementally the internal representation as new information is available. The algorithm is adaptive and tends to privilege new data when building the SV solution. This prevents the old knowledge to nest into the model and eventually become a possible source of misleading information. We tested our approach in the domain of vision-based place recognition. Extensive experiments show that using transferred knowledge clearly pays off in terms of performance and stability of the solution.
BibTeX
@inproceedings{luo2019svm,
  author = {Jie Luo and Andrzej Pronobis and Barbara Caputo},
  title = {{SVM-based Transfer of Visual Knowledge Across Robotic Platforms}},
  booktitle = {Infoscience},
  year = {2019},
  doi = {10.2390/biecoll-icvs2007-120},
}

Loading related content...

Figure