Domitilla Del Vecchio

Other research activities in which I have been involved in include the collective robotics project (CORO), the modeling of the vision system in flies for feedback (Dickinson Lab), the temperature control problem of an industrial tubular reactor at the Ecole Nationale Superieure Des Mines Des Paris with professor Nicolas Petit , and I was in the Caltech team for the synthetic biology competition during Summer 2004.
 

Click image for movie!	Recognition of human motion We approach the problem of recognizing human motion by decomposing human activity into a sequence of meaningful motions, such as reach, step, step over, etc. We believe that in the dynamics of motions there is enough information that allow people to recognize them. As an example consider the picture here on the left: What does it represent? The answer appears clear when the image is animated. Thus, using tools from dynamical systems theory, system identification theory, and pattern recognition, we develop segmentation and classification algorithms that allow to decompose a data stream into the composing atomic motions that we refer to as movemes. We parameterize a moveme by means of the parameters of a fitting linear dynamical system. Our data sets, composed mainly by wrist motions, show that elementary human motions can be successfully distinguished on the basis of a small number of dynamical parameters (2-5). For classification results on real human motion data captured with a computer mouse see Reach-Draw Clusters . For segmentation and classification results on human drawing data see Segmentation-Classification: here the yellow dots are the segmentation points found by the algorithm. If a segment between two yellow dots is blue, it means that the algorithm classifies such a segment as a draw, if a segment is red, the algorithm classifies it as a reach, if a segment is green, the algorithm classifies it as a circle. This work is developed with professors Richard M. Murray and Pietro Perona (Vision Lab)
Click image for movie!	Observability of distributed decision and control systems using partial order theory We consider the problem of estimating the internal state (hidden variables) of a class of decision and control systems characterized by both continuous variables evolution and discrete variables evolution. In particular, we are concerned with decentralized multi-agents systems, such as are found in robot soccer, mobile sensor networks, air-traffic controlled systems, where the continuous variables represent physical quantities such as position and velocity, and discrete variables represent the state of the internal logical system or communications protocol used by the agents to coordinate their actions. The Figure on the left shows a simplified version of the capture the flag game for robots (RoboFlag). The observation problem of interest is to estimate the internal variables that are not measured directly. This is useful to verify the correctness of an observed behavior as is in the case of air-traffic systems, or it could be used to implement a control strategy from the estimated state. This is the case of the RoboFlag Drill example, where each red robot wants to estimate who is the blue robot that is going to tag it. Serious complexity issues emerge when the space of discrete variables is large. To overcome these complexity issues, we develop a novel methodology for designing the estimators, which relies on lattice and partial order theory. The obtained results are promising and clearly suggest that partial order theory is suitable for analyzing and solving problems in systems that show hybrid dynamical behavior. This work is developed with professors Richard M. Murray and Eric Klavins.
	Inference of rules from data: Identification of decision rules in a human controlled system In the course of a day, humans are faced with many choices: Do I step on the brakes? Do I invest in this stock? Do I throw the basketball to Barry or Steve? Some of these choices are very complex, making claims of predicting their outcomes sound far-fetched. Also, deriving these rules from the actions of a group can be very difficult, making human behavior hard to predict. The ability of modeling and predicting human behavior has many beneficial applications. Security systems could be trained to detect abnormal behavior and report it accordingly; the occurrence of traffic accidents and violations could be detected, aiding emergency medical technicians and law enforcement; in strategic games such as soccer or robot soccer, the strategy of a team could be inferred from data. We have developed a preliminary study in which we focus our attention to a particular human controlled system: vehicles at a traffic intersection. There are many rules followed at an intersection. Some are clear-cut, like the laws of traffic and the color of the traffic lights. Others, like when to turn or stop, or when to decelerate or accelerate, are less obvious. A traffic intersection is therefore a good place to test learning algorithms, providing both trivial and nontrivial rules to test an algorithm on. We tested our algorithm on a human-controlled experiment. Using the Multi-Vehicle Lab at Caltech, we ran experiments where human subjects were remotely driving kinematic robots (picture left) through a three-way traffic intersection. The results obtained on our data set are promising. We obtained 15.81% training error and 16.13% test error when training and testing were performed on different subjects (movie). This work is developed with professor Richard M. Murray and Claire Walton (undergraduate student at CalTech). The picture: These are the robots used in some of my experiments. The robots are equipped with a PC 104, a flash disk (256MB) running Slackware 9.0 (Linux kernel), and an orinoco wireless card (802.11b).
	Genetic networks models and dynamics The implementation in living cells of synthetic genetic circuits with an expected behavior provides a way of understanding biological mechanisms, as well as it provides a potential way of building "useful" living circuits. Dynamical systems theory and control theory provide a wide range of tools that can be used in order to design, model, and analyze such circuits. We have been concentrating on gene circuits that can model the circadian clock of living cells. Thus, circuits that show oscillatory behavior have been designed and analyzed (see CDS 270-3 (2004) course report for a brief overview). The circuit on the left represents a possible model for a relaxation oscillator in the A and B protein concentrations. We are currently implementing this circuit in E.Coli in Michael Elowitz Lab.