Jan-Willem van de Meent

Papers

Total Citations

H-Index

About

Jan-Willem van de Meent is a leading researcher at the intersection of deep learning, probabilistic inference, and robotics. His work focuses on enabling machines to learn and generalize from minimal data, particularly in complex, high-dimensional environments. A central theme is the development of structured world models that capture the underlying factors of a scene. His most-cited paper, "Learning Discrete State Abstractions With Deep Variational Inference" (2020, 6 citations), introduces an information bottleneck method for learning approximate bisimulations, allowing agents to compress large state spaces into meaningful representations for efficient decision-making. In robotics, van de Meent has pioneered methods for few-shot imitation learning. His work "One-shot Imitation Learning via Interaction Warping" (2023, 3 citations) demonstrates how a robot can learn a manipulation policy from a single demonstration by inferring 3D object meshes. He also addresses the combinatorial explosion in multi-object environments with "Factored World Models for Zero-Shot Generalization in Robotic Manipulation" (2022, 3 citations). Additionally, his research on "Action Priors for Large Action Spaces in Robotics" (2021, 2 citations) tackles the challenge of learning useful policies without extensive reward shaping or expert demonstrations. Van de Meent’s contributions are shaping a future where robots can learn faster, generalize better, and operate more autonomously in open-ended environments.