Coline Devin

About

Coline Devin is a robotics and machine learning researcher whose work sits at the intersection of robot learning, transfer learning, and visual representation for autonomous systems. Her research addresses one of the field's central challenges: enabling robots to acquire and generalize skills across diverse environments, embodiments, and tasks without prohibitive amounts of real-world data. Among her most influential contributions is her early work on sim-to-real transfer and invariant feature spaces for reinforcement learning (each exceeding 100 citations), which demonstrated how agents can learn from observing morphologically different creatures and bridge the gap between simulation and physical deployment. Her 2018 Grasp2Vec work pioneered self-supervised object-centric representations for robotic manipulation, while her object-centric approach to autonomous driving (103 citations) improved interpretability and robustness in visuomotor policies. Devin also contributed to the landmark Open X-Embodiment project (119 citations), a large-scale collaborative effort producing general-purpose robotic foundation models trained across diverse datasets — a milestone toward universally capable robots. Her modular neural network policy work further advanced multi-task and multi-robot transfer. Collectively, her research has meaningfully shaped how the robotics community thinks about generalizable, data-efficient robot learning.

Research Focus

Key Achievements

Top Papers

1. 1
   Open X-Embodiment: Robotic Learning Datasets and RT-X Models : Open X-Embodiment Collaboration⁰

   119 citations · 2024
2. 2
   Learning Invariant Feature Spaces to Transfer Skills with Reinforcement\n Learning

   117 citations · 2017
3. 3
   Deep Object-Centric Policies for Autonomous Driving

   103 citations · 2019
4. 4
   Open X-Embodiment: Robotic Learning Datasets and RT-X Models

   101 citations · 2023
5. 5
   Adapting Deep Visuomotor Representations with Weak Pairwise Constraints

   81 citations · 2020
6. 6
   Grasp2Vec: Learning Object Representations from Self-Supervised Grasping

   74 citations · 2018
7. 7
   Towards Adapting Deep Visuomotor Representations from Simulated to Real Environments.

   61 citations · 2015
8. 8
   Learning modular neural network policies for multi-task and multi-robot transfer

   38 citations · 2017
9. 9
   SMiRL: Surprise Minimizing RL in Dynamic Environments

   17 citations · 2019
10. 10
    Beyond Pick-and-Place: Tackling Robotic Stacking of Diverse Shapes

    16 citations · 2021

Key Collaborators

Contact & Links