Self-supervised Mobility Assessment from Unsupervised Proprioceptive Feature Learning on Simulated Environment
Abstract
This work leverages open source tools to create a virtual environment and vehicle that compare to real world conditions in order to tests, train and evaluate different machine learning approaches. In particular, the combination of a Convolutional and a Long Short-Term Memory Recurrent Neural Networks under a Self-supervised learning strategy to avoid extensive labeling and enable the learning of new patterns and terrain, particular that affect or restrict mobility. In the self-supervised structure, the LSTM produces continuous results to provide to the CNN. Since the camera is front looking and the LSTM only classifies once it physically reaches the surface the strategy requires a network configuration that allows making both types of data compatible to create a common map. The retraining CNN layers get passed to work online through a Frozen feature extractor. Running each model through the map and normalizing the results to make them comparable shows different levels of accuracy against validation data using weakly and fully labeled data on both networks separately as well as the combination in the proposed Self Supervised structure. This shows results in the range of 85.93% against validation data for the individual networks and a much more promising 94.61% over weakly labeled data. A better results on the combined approached was excepted and hypothesized since it allows the networks to use different source representing the same class of data.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 01, 2022
- Accession Number
- AD1172808
Entities
People
- Andres Carcamo
- Eric Mark
- Ian Hughes
- Jose L. Verdugo
- Jose M. Larenas
- Paramsothy Jayakumar