Alexander Pashevich

I am a Ph.D. specializing in deep learning, robotics, computer vision, natural language processing, and time-series forecasting. During my doctoral studies at Inria, I pioneered robot learning research within the team and developed methods that enabled robots to acquire visually and language-guided behaviors. My work, published in top-tier venues, demonstrated the advantages of learning-based approaches compared to classical control algorithms. After transitioning to industry, I have been developing deep learning models for credit adjudication and fraud detection, driving more than $100M+ in annual business impact. I am passionate about translating cutting-edge AI research into real-world solutions.

Email  /  Google Scholar  /  LinkedIn  /  Github  /  CV

My research interests include robotics, reinforcement and imitation learning, simulation-to-reality transfer, natural language processing, and vision-and-language navigation. More recently, I have been working on time-series forecasting research. During my Ph.D. studies, I was adviced by Cordelia Schmid and co-adviced by Ivan Laptev. I also had a pleasure to work with Chen Sun during my internship at Google Research.

A simple, inexpensive, theoretically motivated regularization term to enhance the robustness of deep time-index models for time-series forecasting.

A multimodal transformer-based architecture for vision-and-language navigation (VLN) improving results on a challenging task by 74%.

An approach learning to build shapes by disassembly that combines learning in low-dimensional state space and high-dimensional observation space.

A reinforcement learning approach to task planning that learns to combine primitive skills learned from demonstrations.

An approach for transferring policies learned in simulation to real-world robots by exploiting parallel computations.

A hierarchical reinforcement learning approach for learning from sparse rewards.

An algorithm for unsupervised extraction of piecewise planar models from depth data using constrained Gaussian mixture models.

website template credit