Maintained by: tpaine@mit.edu Get PDF
Ensemble methods for online estimation of dynamical models for marine vehicles
In this (side) project we investigated using an ensemble of three different methods for online system identification for dynamical models of marine vehicle motion. The three methods were a contracting recurrent neural network (RNN), adaptive identifier (AID) and recursive least squares (RLS). For the RNN, we designed and implemented a new term in the loss function to encourage learning a stable RNN in the sense of contraction stability. The ensemble was implemented onboard 8 Heron USVs and ran as a background process in the summer and fall of 2022 for a total of more than 30 hrs of run time. Overall the results were promising, and the best result was that the ensemble on each of the 8 vehicles learned a unique model, and that model was uniquely optimized for that particular vehicle. Given the initial success, we aim to apply these methods towards fault identification and mitigation in addition to using the model to better inform autonomous decision making. More details can be found at
Figure 1.1: Eight Heron Unmanned Surface Vehicles (USVs) made by Clearpath Robotics were used for field testing. Inset: Block diagram of online ensemble estimation using adaptive identifier (AID), recursive least squares (RLS), and recurrent neural network (RNN) approaches.
Status: | Ongoing since January 2023 |
Sponsor(s): | In-House |
People: | Tyler Paine, Mike Benjamin |
Robots: | https://oceanai.mit.edu/pavlab/robots/herons |
Software: | MOOS-IvP public codebase, MOOS-IvP-Pavlab codebase |
Recent Publications
2023 (1 item)
- Tyler Paine, Michael Benjamin, An Ensemble of Online Estimation Methods for One Degree-of-freedom models of Unmanned Surface Vehicles; Applied Theory and Preliminary Field Results with Eight Vehicles, International Conference on Intelligent Robots and Systems (IROS), October, 2023. (bibtex)
Document Maintained by: tpaine@mit.edu
Page built from LaTeX source using texwiki, developed at MIT. Errata to issues@moos-ivp.org. Get PDF