Modelling Robot Navigation Recovery Policies from Non-Expert Users’ Demonstrations In-The-Wild

Learning effective robot policies from non-expert users in real-world environments is a significant challenge, as demonstration data is often small, unstructured and suboptimal. This paper presents a novel framework for modelling navigation recovery policies directly from such “in-the-wild” human interactions. Our framework is composed of (a) a Policy Encoder Model (PEM) to transform the high-dimensional input data into a latent representation for learning; (b) a Latent Dynamic Model (LDM) to represent transitions in the latent space; and (c) a Gaussian Process Policy (GPP) that learns to map latent states to recovery actions, guided by an unsupervised goal discovery module. We evaluated our system on a dataset of recovery trajectories generated by visitors helping a tour guide robot in a public museum. On our unseen test trajectories, our model achieves 100% success rate in our end-to-end latent space navigation. The learned policies can reproduce demonstrations with high spatial fidelity (Euclidean Reconstruction R² Score of 0.985) but also generalise to find novel recovery paths. This work demonstrates a promising direction to learning robust and generalisable robot behaviours from non-expert data, providing a pathway to more adaptable autonomous systems.

 @inproceedings{faris2025modelling,
  title = {Modelling Robot Navigation Recovery Policies from Non-Expert Users' Demonstrations In-The-Wild},
  author = {Faris, Ahmad and Kucukyilmaz, Ayse and Polydoros, Athanasios and Del Duchetto, Francesco},
  booktitle = {2025 International Conference on Robotic Computing and Communication (RoboticCC)},
  pages = {15--22},
  year = {2025},
  month = dec,
  doi = {10.1109/RoboticCC68732.2025.00014},
  organization = {IEEE},
  public = {yes},
  pdflink1 = {https://doi.org/10.1109/RoboticCC68732.2025.00014}
}