Abstract

In autonomous navigation, there are numerous applications in which audio plays a crucial role as an essential source of information. This study investigates the efficacy of employing transfer learning for optimal path planning through multiple sound-emitting destinations. This problem is challenging due to the intricate feature extraction from mixed audio signals and combinatorial complexity inherent in multi-destination path planning. Expanding beyond the current reinforcement learning study for the single sound source scenario, we present a multi-targeted formulation and explore how effectively fine- tuning a pre-trained agent adapts its performance to the multi- sound source scenario. We provide a rigorous evaluation of our proposed multi-source approach on the widely adopted Matterport3D dataset to showcase its effectiveness. The test results underscore a notable acceleration in the training process by more than one order of magnitude.

Hypotheses

H.1 Pre-trained Audio-Visual Navigation Policy in single-target scenarios transfers to multi-target scenarios. In particular, pre-training an agent on a single-target task expedites the convergence process in multi-target scenarios and leads to optimal performance in a fraction of the training updates.
H.2 Pre-trained Audio-Visual Navigation Policy in single-target scenarios transfers to random number of target scenarios. Specifically, a pre-trained agent on the single target task generalizes to an arbitrary number of destinations in an effective manner.

Results

Comparision of training cost with/without transfer learning

Test trajectories

Videos

Acknowledgements

We thank the Research Computing department of University of South Carolina for providing compute support.

The website template was borrowed from Michaël Gharbi.