Andrew Jaegle

About

Andrew Jaegle’s research lies at the intersection of machine learning, robotics, and decision-making, with a focus on learning latent dynamics, modular reinforcement learning, and active data acquisition. His work addresses fundamental challenges in enabling intelligent systems to reason about complex, high-dimensional physical environments—such as those encountered in robotics and autonomous driving—where algorithms must infer underlying dynamics from raw sensory inputs like images. A key contribution is his benchmarking of models for learning latent dynamics, which provides a rigorous framework for evaluating how different priors affect performance in these settings. While his most-cited paper, “Which priors matter? Benchmarking models for learning latent dynamics” (2021, 4 citations), establishes a foundation for understanding representation learning in dynamic systems, Jaegle also explores modular reinforcement learning to bridge abstract objectives with concrete motor control in physically embedded tasks like 3D Sokoban. His more recent work introduces the challenging task of active acquisition for multimodal temporal data, where agents must strategically acquire costly input features at test time—a problem with direct relevance to real-world sensor management. Through these contributions, Jaegle advances the frontier of embodied intelligence and efficient decision-making under uncertainty.

Research Focus

Key Achievements

Top Papers

1. 1
   Which priors matter? Benchmarking models for learning latent dynamics

   4 citations · 2021
2. 2
   Beyond Tabula-Rasa: a Modular Reinforcement Learning Approach for Physically Embedded 3D Sokoban

   2 citations · 2020
3. 3
   Active Acquisition for Multimodal Temporal Data: A Challenging Decision-Making Task

   2 citations · 2022

Key Collaborators

Contact & Links