Robotics and Automation Lab Internet Delay Prediction for Teleoperation Imad Elhajj September 4, 2002
Robotics and Automation Lab Applications Tuning Control Parameters (gains) Control of Quality of Service (reduce resolution) Tuning Communication Parameters (discarding packets, reducing frequency)
Robotics and Automation Lab Assumptions/Limitations Delay can be predicted with finite dimension expression Delay is correlated (function of previous values) Delay is the same in both directions
Robotics and Automation Lab A Delay Prediction Approach for Teleoperation over the Internet Tissaphern Mirfakhrai and Shahram Payandeh Simon Fraser University BC Canada
Robotics and Automation Lab Algorithm Train the system to find appropriate gain for each delay Train the delay model Predict delay based on model Look up best gain for that delay from table
Robotics and Automation Lab Models for delay Moving Average –Low pass filter –Future value is average of its past values Autoregressive –Weighed average –Includes white noise
Robotics and Automation Lab Autoregressive X[n] is the signal modeled N is order of model (design parameter) W[n] white noise with autocorrelation
Robotics and Automation Lab Error Versus Order N
Robotics and Automation Lab Predicted Delay Versus Actual
Robotics and Automation Lab Errors at Different Delays for Different Parameter Values
Robotics and Automation Lab Simulation Results
Robotics and Automation Lab End-to-End Delay Boundary Prediction Using Maximum Entropy Principle (MEP) for Internet-Based Teleoperation Peter Liu, Max Meng University of Alberta, Canada Xiufen Ye Harbin engineering University, China Jason Gu Dalhousie University, Canada
Robotics and Automation Lab Traditional autoregressive and moving average models are not suitable for fast changing delays Roundtrip time delay (RTT) Roundtrip timeout (RTO) evolves –RTO too small unnecessary retransmission –RTO too large bandwidth wasted
Robotics and Automation Lab Traditional TCP R i actual RTT R` i predicted RTT ` RTT variation prediction
Robotics and Automation Lab The model
Robotics and Automation Lab Reflex Condition The coefficients of the model still posses validity when used to calculate R n. The reflex condition is used for this purpose,
Robotics and Automation Lab Maximum Entropy Principle We know a ni are positive and their sum is 1 a ni are considered as a set of probability distributions Reflex condition is contraint on set of distributions There are infinite distributions satisfying the reflex condition Maximum entropy principle is most unbiased to choose a ni
Robotics and Automation Lab MEP Shannon entropy, is maximal under the reflex condition.
Robotics and Automation Lab Maximum Entropy Principle Where n is the solution of the equation,
Robotics and Automation Lab
Conclusions Predicting delay has benefits Delay prediction does not solve all the problems of teleoperation Existing methods are not highly accurate (up to %50 error)
Robotics and Automation Lab Internet Control of Personal Robot Between KAIST and UC Davis Kuk-Hyun Han, Yong Kim, Jong-Hwan Kim Korea Advanced Institute of Science and Technology Steve Hsia University of California, Davis
Robotics and Automation Lab Remote control of personal robot Audio and Video feedback (netmeeting) Three modes of control –Direct –Supervisory –Job Scheduling
Robotics and Automation Lab