-
Notifications
You must be signed in to change notification settings - Fork 1
HongminWu/HDPHMM_HDPSLDS
Folders and files
Name | Name | Last commit message | Last commit date | |
---|---|---|---|---|
Repository files navigation
@ Hongmin Wu 1. This is based on software originally written by Emily B. Fox and adapted for using in the fault detection and recognition in robot contract task. if you want to get further informations, please don't hesitate to contact me in Github/Email. 2. The main script to run is "demos" folder, there are two phase to accomplish the fault detection and recognition, first, demos/demo_training: the script denotes that training the specific state by nonparametric hmm and inference the unknown parameters second, demos/demo_testing: we use the log-likelihood of observations to calssify the success/failure by calculate the p(y_{i}|trained_HMM_model) finally, global_variables : the global parameters should be defined before run any demos, if you want to create a struct that has new observations of each of your tasks, you can following this file to add appropriate parameters. 3. Why we are crazy about this research? this work investigates the use of Bayesian nonparametric techniques that define a prior on Markov Switching processes with an unbounded number of states for the state estimation of contact tasks. In particular, we explore the use of a sticky Hierarchical Dirichlet stochastic process prior to learn a number of Markov Jump Linear System models for wrench signatures in the contact tasks. We believe the approach will lead to strong generalizations that discover and model the underlying robot state in robust, accurate, and computationally efficient ways. Previous state estimation techniques for contact tasks have often used Markov Models or supervised learning methods. Both of these methods have primarily been parametric, implying that the number of classes or states is defined a priori and fixed through the entire task. This is a limiting assumption as it lacks modelling power to properly segment the data into natural clusters that represent task’s state. Furthermore, Markov Models which are a type of Markov Switching Process, are limited in their ability to model complex dynamical behaviors due to their assumption that observations are conditionally independent given the state. Instead, this work considers more powerful modelling Markov Switching Processes; namely, those in the class of Markov Jump Linear Systems (MJLS). These systems more powerfully describe rich dynamical phenomena through their conditionally linear modes. In particular, the switching vector autoregressive and the switching linear dynamical system methods will be used. Key scientific issues to be solved lie in the determination of dynamic parameters for the MJLS methods and the determination of more informative prior distributions and their corresponding hyper-parameters. These methods will be tested on contact tasks that consist of multiple contact sub-tasks including pushing, pulling, alignments, insertions, mating, sliding, amongst others. We expect to integrate a real-time open source framework that will exploit MJLS system and significantly increase the robustness, accuracy, and computational efficiency of state-of-the art approaches in robot contact task state estimation. We also expect to compile and open to the public a library of robot primitive actions useful in contact state estimation. A wide array of robotic domains will benefit from the ability to confirm commanded actions including: collaborative manufacturing robots, service robots, as well as areas like high-level planning and programming by demonstration. Better state estimates will also derive in error reduction and increased efficiency across all robot domains. 4. we have some datesets of HIRO robot snap assembly can be download at: http://www.juanrojas.net/
About
HDPHMM_HDPSLDS
Resources
Stars
Watchers
Forks
Releases
No releases published
Packages 0
No packages published