Online Slip Prediction for Mobile Robots 16831 Project Proposal Neal Seegmiller, Chris Skonieczny
Prior Work Rogers-Marcovitz , Forrest. “ On-line Mobile Robotic Dynamic Modeling using Integrated Perturbative Dynamics ,” Master's Thesis, tech. report CMU -RI-TR-10- 15, Robotics Institute, Carnegie Mellon University, May, 2010 • Learn dynamic model for a mobile robot that accounts for wheel slip. • “Integrated”: • Most approaches are parameterized: • Vehicle Ground Model Identification Project. Future goals include perception, learning different models for multiple terrain types
16831 Project Objectives OFFLINE (assume constant speed and curvature commands) Gaussian Process Regression Bayes Linear Regression Least Square Regression ONLINE (transient commands allowed) Extended Kalman Filter Objectives: 1. Make online results better match offline batch results 2. Accurately quantify uncertainty
Evaluation metrics • Things quickly forgotten cannot be said to have truly been learned • Two important metrics that capture this: – Proximity of online solution to an offline approach that treats all data points equally – Stability of parameter estimates
Potential online parameter learning approaches • Online Bayes Linear Regression • Re-tuned Extended Kalman Filter – Reduce high weighting of recent data relative to older data • Markov Random Field x i – State: slip for a bin of commanded z i speed and curvature – Measurement: difference between measured and predicted slip • Other suggestions?
Recommend
More recommend
