Research Interests


I'm interested in the research areas in between machine learning and artificial intelligence that will finally narrow the gap between robotics and the real world. I'm specifically talking about the areas that will contribute to create the algorithms to learn models from noisy sensor data to do planning and control under uncertainty. Such task could be accomplished using different approaches like the reinforcement learning perspective working under the assumption of a reward function or creating probabilistic models.

 


Research Background


  • Research Programmer in the Auton Lab at the Robotics Institute -CMU   (2009 July - Current)     
  • Research Visitor AutonLab-Select Lab(2009 Spring)
  • Summer Scholar Program - Machine Learning Department - CMU (2008 Summer)    
  • Researcher Universidad Miltar Nueva Granada (2008) 
  • Young researcher Universidad Militar Nueva Granada (2007)
  • Research assistant Universidad Militar Nueva Granada (2006)


Next a short description of the most important research projects where I have been working:

CMU audio training and clasiffication API : API for training and scoring of Hidden Markov Models using the VSTACS algorithm.  This project was a joint effort between Mobile Fusion Inc and the Auton Lab. Our goal was the classification of sounds in between 6 labels defined as human, animal, ground vehicles, aerial vehicles, explosions and background. The results are included in the paper Automatic State Discovery ofr unstructured audio scene analysis in ICASSP 2010, it will be available after May 2010.

STACS (modification): I basically modify this intrincated, hardcoded and long program (and without documentation) basically to add more functionality as well as implement a new stoping criteria, the results are pending for publication. Elsewhere the previous version and the paper that explains how the algorithm works can be found in Sajid's Siddiqi web site.

Trogdor (Refurbishing): This project was done during the summer of 2008, was my first time at CMU as well as my first time with Linux, Matlab and in general lots of things. Despite the lack of documentation and the short time that I had to accomplish my goals, I succesfully get working the Obot d100 - Trogdor. Also as part of my job I add the Bumblebee2 camera to the Carmen Robotics Plataform. For that I had to get the camera working first in linux (the provider doesn't support any longer linux) and then within CARMEN (The Carnegie Mellon Robotics Platform). Also I implemented Structure From Motion on a stream of images using Scale Invariant Features. All this work was done under supervision of  Dr.Geoff Gordon.



EASY Bot : This was my second robot. Is equiped with an Eyecon mk5, as well as a system for gas sensing. This robot will be able to find odour sources of amonia, ethanol and in general volatitle organic compounds. Unfortunately I couldn't continue working in that project since I got a full time position as a Research Programmer at CMU.  This was the extension of the research that I did during my bachelor (Odour recognition and now localisation). With the robot was also developed a vision odometry based system with an external camera in the roof, the location as well as the orientation of the robot was logged all the time, keeping track of gas concentrations at any time. This all done trying to figure out the best strategy for the odour source localization, before coming back to CMU a couple of interesting experiments were done also I got a gas concentration map that give me an idea on how was the behaviour of the gas in the lab. The next steps defined for the project were basically the implementation of Moth Behaviour and Fruit Fly strategies. The first one is well known, but in this specific case unfortunately the robot doesn't count with sensor to detect the direction of wind. The second strategy is much more simple and doesn't rely on wind direction. Also it seems accurate and simple enough to be carried on by the robot. 

robot, odour source localization

The Next projects were done during my undergraduate courses

System for Odour recognition : First serious research project using Neural Networks. This system is able to recognize between different volatile organic componounds at different humidity, concentration and temperature levels. Within this project was created the program ICARUS (r) and ICARUS SE (r). The last is able to create a neural network of any given size, allows the user to train the net, test or save for later training, also allows for different stoping criterias as the real media squeared measured using a validation set of patterns. The system was composed by a smell chamber, a sensor matrix,  an acquisition card designed by me and the ICARUS. Three publications, two sotware registers and two prototypes  were derivated from this project please check the publications link.

Obi-1 : This one was my first robot. Was able to pick up a pingpong ball, take it upstairs and deposit in a given site autonomously. To acomplish this task was used basic computer vision and for the autonomous navigation fussy logic. Unfortunately and for my regret there's nothing else than  a couple of bad photographs and the software left and of course the robot. Unfortunately there's no video of it working.

Hovercraft(2005) : This was my first research project,  I and my team made it from scratch, there were no information, no books, no equations, just a bunch of ideas and a couple of principles on the internet to follow. The only thing that we had left was experimentation, that's how we made five prototypes each one proving different principles of functioning , each one proving some efficiency and problems. The final version was a summary of all the prototypes, it proved to be very efficient and good enough to win the challenge. 

Hovercraft

Videos 
First test 
Second test
Final presentation (on brick)
Final presentation (on water)

Hopfield Networks(2004)
This is just an small program that reconstructs a given pattern, in this case is just for numbers. Check it out here. This program uses Hopfield Networks using an old algorithm which has the problem of generation of spurious states more recen algorithms avoid that problem and also the onerous calculation of a huge matrix.

Manipulator (2004)
This was an early version of program that calculates the direct and inverse kinematic of a Manipulator with 3 angular joints. Used to work fine only in PCs with a GPU in my case was a geforce fx 6200. 

Link to the .exe

.

 

 

 

  1. Ramos J. A. Siddiqi, S., Gordon G., Dubrawski A., and Sharmma A.,  "Automatic state discovery for unstructured audio scene analysis". To be published in ICASSP 2010

 

  1. Ramos J. A. and Vargas W. L. “Pattern recognition from a sensor array using neural networks”. Proceedings of 23rd ISPE International Conference on Cad/Cam Robotics and factories of the future. ISBN 978-958-978-597-3 (2007). (pdf)

 

  1. Ramos J. A. and Vargas W. L. “Reconocimiento de patrones en un arreglo sensorico usando redes neuronales”. Ciencia e Ingeniería Neogranadina. ISSN 01248170 Vol. 17 No.1 (2007).(pdf)

 

  1. Ramos J. A. and Vargas W. L. “Icarus, Software para reconocimiento de patrones de arreglos sensoricos”. Proceedings of 3rd IEEE Colombian Workshop on Robotics and Automation. ISBN 978-958-978-658-1 (2007).(pdf)

 

Personal Information

Julian Andres Ramos Rojas

NSH 3122 - Robotics Institute - Carnegie Mellon University - 5000 forbes avenue, Pittsburgh PA (15213) - United States

Phone 412-268-1238

e-mail: ingenia at andrew dot cmu dot edu  

ing dot julianr at gmail dot com

CV Last update (December 2009)


 

Julian Andres Ramos Rojas