I am currently working as a Data Scientist at Talan working as a consultant for the RATP Group. I was a former postdoctoral researcher in Automatic Control at UGE formely IFSTTAR. My recent research interests include Networked Control Systems (NCS), Cyberphysical Systems (CPS), Intelligent Transportation Systems (ITS) and Artificial Intelligence (AI). Previously, I did my PhD at INRIA/CNRS with the NeCS team developing Short-term Forecasting and estimation techniques for large scale traffic networks. During my last research experience, I worked at the LICIT team working on control techniques for connected and automated vehicles (CAV)s.
I am passioned about the Data Science community and the recent work developed around reproducible research. Fan of #rstats and python.
PhD Automatic Control, 2018
Université Grenoble Alpes
M.E. Electronic Engineering, 2012
Pontifical Xavierian University
B.E. Electronic Engineering, 2008
Pontifical Xavierian University
Manipulation, DataViz, Modeling
Predictive, Non-linear, Gurobi, CVX
Hardware in the Loop / Control toolbox
Hardware Oriented Development
PyViz, Pandas, Keras
Development of estimation and prediction techniques for traffic systems:
The growing number of connectivity services in transportation along with a deluge of data collected from different sources, make new control approaches possible to reduce traffic congestion inside urban areas by leveraging V2I (vehicle to infrastructure) communication. Our approach aims at reducing congestion via a (real-time) dynamic rerouting of a fraction of vehicles. The infrastructure broadcasts aggregated speed information for predefined zones, then a cooperative strategy allows the qualification of each zone in terms of traffic performance. This information is then asynchronously used by each of the vehicle agents to determine new paths and achieve their destination. For the sake of scalability, our solution is distributed, using multi-agent interaction and computation based on a framework
This paper presents a new concept for operating traffic management at a sizeable urban scale. The overall principle is to partition the network into multiple regions, where traffic conditions should remain optimal. Instead of monitoring the inflow, like for perimeter control, deviations in regional mean speed compared to the reference are transformed into avoidance levels by a centralized controller. That information is broadcasted to vehicles through a public avoidance map. The onboard navigation systems then interpret this map to deliver individual route guidance according to the avoidance levels. In that way, the concept tends to distribute vehicles in the network optimally while preserving privacy. In addition to the controller parameters itself, three other critical parameters define the system: the safety distance, the controller time horizon, and the region sizes. Thorough sensitivity analysis leads to the optimal setting. The concept is proven effective using microsimulation for both a Manhattan and a realistic network. The total travel times is improved by about 15% when traffic is severely congested compared to the uncontrolled case. In the meantime, the mean individual travel distance increase for rerouted vehicles is kept below 10%. Those results have been obtained with a simple decentralized reactive control framework, i.e., an individual proportional feedback system independently governs each region. More advanced formulations introducing cooperation between the regions are also tested. In a nutshell, this paper is a proof of concept for a new control system that appears both practical and valuable to alleviate congestion in urban areas
Lane change maneuvers are main causes of traffic turbulence at highway bottlenecks. We propose a dynamic game framework to derive the system optimum strategy for a network of cooperative vehicles interacting at a merging bottleneck. Cooperative vehicles on the highway mainline seek for optimal strategies (i.e. whether and when to perform courtesy lane change to facilitate the merging vehicle) to minimize their cost, while taking into account potential future interactions at the merging section while minimizing the distance travelled on the acceleration lane. An optimal strategy is found by minimizing the joint cost of all interacting vehicles while respecting behavioral and physical constraints. Numerical examples show the feasibility of the approach in capturing the nature of conflict and cooperation during the merging process and demonstrate the benefits of sharing information and cooperative control for connected automated vehicles.
A key question about cooperative vehicle longitudinal control is reactivity, which determines the future of road safety, and capacity. In adaptive cruise control (ACC), the controller adapts the speed of the vehicle to its immediate leader’s speed whereas, in the cooperative version (CACC), connectivity between the platoon equipped vehicles reduces their response times. The USDoT Cooperative Automated Research Mobility Applications (CARMA) platform provides data for platooning experiments involving ACC and CACC vehicles. We measure ACC response times (mean = 2.78 seconds) larger than for human-driven cars. We study response times inside CACC platoons showing that connectivity is not always effective.
Truck platooning has attracted substantial attention due to its pronounced benefits in saving energy and promising business model in freight transportation. However, one prominent challenge for the successful implementation of truck platooning is the safe and efficient interaction with surrounding traffic, especially at network discontinuities where mandatory lane changes may lead to the decoupling of truck platoons. This contribution puts forward an efficient method for splitting a platoon of vehicles near network merges. A model-based bi-level control strategy is proposed. A supervisory tactical strategy based on a first-order car-following model with bounded acceleration is designed to maximize the flow at merge discontinuities. The decisions taken at this level include optimal vehicle order after the merge, new equilibrium gaps of automated trucks at the merging point, and anticipation horizon that the platoon members start to track the new equilibrium gaps. The lower-level operational layer uses a third-order longitudinal dynamics model to compute the optimal truck accelerations so that new equilibrium gaps have been created when merging vehicles start to change lane and the transient maneuvers are efficient, safe and comfortable. The tactical decisions are derived from an analytic car-following model and the operational accelerations are controlled via model predictive control with guaranteed stability. Simulation experiments are provided in order to test the feasibility and demonstrate the performance and robustness of the proposed strategy.