As a research scientist at the Toyota Research Institute, I bring my expertise in machine learning with a doctorate from Paris Saclay University. My current research interests include developing embodied machine learning solutions for sensory-motor systems and exploring the frontiers of robotics with human interactions. Through my work, I am dedicated to advancing state-of-the-art ML research for robotics and driving meaningful innovation.
PhD in Machine Learning, 2021
MEng in Scientific Computing, 2015
Robust planning in interactive scenarios requires predicting the uncertain future to make risk-aware decisions. Unfortunately, due to long-tail safety-critical events, the risk is often under-estimated by finite-sampling approximations of probabilistic motion forecasts. This can lead to overconfident and unsafe robot behavior, even with robust planners. Instead of assuming full prediction coverage that robust planners require, we propose to make prediction itself risk-aware. We introduce a new prediction objective to learn a risk-biased distribution over trajectories, so that risk evaluation simplifies to an expected cost estimation under this biased distribution. This reduces the sample complexity of the risk estimation during online planning, which is needed for safe real-time performance. Evaluation results in a didactic simulation environment and on a real-world dataset demonstrate the effectiveness of our approach. The code and a demo are available.
This paper presents a novel vehicle motion forecasting method based on multi-head attention. It produces joint forecasts for all vehicles on a road scene as sequences of multi-modal probability density functions of their positions. Its architecture uses multi-head attention to account for complete interactions between all vehicles, and long short-term memory layers for encoding and forecasting. It relies solely on vehicle position tracks, does not need maneuver definitions, and does not represent the scene with a spatial grid. This allows it to be more versatile than similar model while combining many forecasting capabilities, namely joint forecast with interactions, uncertainty estimation, and multi-modality. The resulting prediction likelihood outperforms state-of-the-art models on the same dataset.
We study the design of learning architectures for behavioural planning in a dense traffic setting. Such architectures should deal with a varying number of nearby vehicles, be invariant to the ordering chosen to describe them, while staying accurate and compact. We observe that the two most popular representations in the literature do not fit these criteria, and perform badly on an complex negotiation task. We propose an attention-based architecture that satisfies all these properties and explicitly accounts for the existing interactions between the traffic participants. We show that this architecture leads to significant performance gains, and is able to capture interactions patterns that can be visualised and qualitatively interpreted. Videos and code are available at this https URL.
In the recent vehicle trajectory prediction literature, the most common baselines are briefly introduced without the necessary information to reproduce it. In this article we produce reproducible vehicle prediction results from simple models. For that purpose, the process is explicit, and the code is available. Those baseline models are a constant velocity model and a single-vehicle prediction model. They are applied on the NGSIM US-101 and I-80 datasets using only relative positions. Thus, the process can be reproduced with any database containing tracking of vehicle positions. The evaluation reports Root Mean Squared Error (RMSE), Final Displacement Error (FDE), Negative Log-Likelihood (NLL), and Miss Rate (MR). The NLL estimation needs a careful definition because several formulations that differ from the mathematical definition are used in other works. This article is meant to be used along with the published code to establish baselines for further work. An extension is proposed to replace the constant velocity assumption with a learned model using a recurrent neural network. This brings good improvements in accuracy and uncertainty estimation and opens possibilities for both complex and interpretable models.
Nanotemplates derived from the self‐assembly of AB‐type block copolymers provide an elegant route to achieve well‐defined metallic dot arrays, even if the variety of pattern symmetries is restricted due to the limited number of structures offered by microphase separated diblock copolymers. A strategy that relies on the use of complex network structures accessible through the self‐assembly of linear ABC‐type terpolymers is presented for the formation of metallic nanodots arrays with “outside‐the‐box” symmetries. Patterned templates formed by the cubic Q214 and orthorhombic O70 network structures are used as excellent platforms to build well‐ordered gold nanodot arrays with unique p3m1 and p2 symmetries, respectively. A simple yet efficient blending strategy is used to tune the critical dimensions of the p3m1 pattern while laterally ordered gold nanodot arrays are also demonstrated through a directed self‐assembly approach. Such highly ordered gold nanodots with tunable particle dimensions and array periods, enabling the control of their plasmonic responses, are attractive probes for biological imaging.
This work reports the emergence of metachronal waves in cilia arrays immersed in a two-fluid environment using a coupled lattice Boltzmann - Immersed Boundary method. The periciliary layer (PCL) is confined between the wall and the mucus layer. Its depth is chosen in such a way that the tips of the cilia can penetrate the mucus layer. The cilia are initially set in a random state but quickly synchronize with their immediate neighbors with a phase shift giving birth to sympleptic or antipleptic metachronal waves, depending on the strength of the fluid retroaction onto the cilia. Antiplectic waves are found to be the most efficient to transport mucus compared to other random or synchronised cilia motions.
Cilia are flexible elongated whip-like structures which are ubiquitous in nature. Indeed, the collective beating of arrays of thousands of cilia can transport fluid (mucus in airways) or induce locomotion on microorganisms swimming in water. From a purely hydrodynamical point of view, cilia do not beat randomly, but rather generate typical metachronal waves at their surface. In this work, we study the self-organization of the beating motion of large fields of beating cilia in a two-component flow environment, made of water and a much more viscous fluid. The numerical solver is based on an immersed boundary-lattice Boltzmann method in the context of single- and multi-component fluid flows, and in the presence of fixed or moving solid boundaries. The solver has been validated in previous studies. Various parameters are varied, such as length, spacing and phase motion of individual cilia. The energetic performances of different kind of waves are studied to understand the emergence of antiplectic metachronal waves, commonly observed in nature. It is found that a purely hydrodynamical coupling between fluid and cilia can explain the onset of metachronal waves in cilia arrays, and that these waves are maximizing a performance ratio.