Interactive Terrain Affordance Learning via VAE Query Selection & Data Manipulation

Jordan Sinclair, Brian Reily, Christopher M. Reardon

Confirmation: I have read and agree with the workshop's policy on behalf of myself and my co-authors.

Keywords: Terrain Affordance Learning, VAE Query Selection

Abstract:

Terrain preference learning from trajectory
queries allows complex reward structures to be obtained for
robot navigation without the need for manual specification.
However, traditional offline preference learning approaches suffer
from ambiguous trajectory pairs stemming from inadequacy
in the initial dataset, which causes longer learning times and
may lead to less accurate results. Several approaches have been
introduced to tackle this common problem including creating
preference learning models robust to volatility in weights from
ambiguous choices, enhancing the query selection process towards
mitigating dubious trajectory choices, and modifying the
original dataset with highly variant samples. Inspired by recent
work in the application of deep learning towards improving
query selection, this paper introduces a joint dataset and query
optimization procedure utilizing variational autoencoders. Our
efforts leverage both the encoder and decoder models to identify
underrepresented terrain types and supplement the trajectory
set with targeted samples created using the decoder. We jointly
optimize a clustered latent space towards creating balanced
clusters that can be used to obtain diverse trajectory pairs.

Submission Number: 10