Kelsey Allen

Kelsey Allen is a cognitive scientist and roboticist who studies how people and machines use tools and build models of the physical world. After a physics degree at the University of British Columbia in 2014, she completed a PhD in Brain and Cognitive Sciences at MIT in 2021 under Josh Tenenbaum, and her thesis won a Glushko Dissertation Prize in 2022. A 2018 paper on tool-use planning with differentiable physics earned a Robotics: Science and Systems best paper award, her 2023 FIGNet paper built learned physics simulators from graph networks, and she worked for four years from 2020 as a research scientist at Google DeepMind. Since 2025, she is an assistant professor at the University of British Columbia in computer science and psychology, a faculty member at the Vector Institute and a Canada CIFAR AI Chair, and in May 2026 CIFAR named her a Global Scholar for 2026–2028.

Now at: University of British Columbia and Vector Institute
Badges: Bridge
Boomerang

Timeline

2008–2014

Physics at UBC

She begins undergraduate studies in physics at the University of British Columbia in 2008 and earns her BSc in 2014. In 2014, she also contributes analyses to an ATLAS Collaboration search for high-mass dilepton resonances.
2018

Best paper on tool-use planning

In 2018, she coauthors "Differentiable physics and stable modes for tool-use and manipulation planning" with Marc Toussaint, Kevin Smith and Josh Tenenbaum at Robotics: Science and Systems. The paper combines task and motion planning with differentiable physics constraints to solve physical puzzles that involve tool use, and it earns a best paper award.
2020

Trial-and-error tool use in PNAS

In 2020, she publishes "Rapid trial-and-error learning with simulation supports flexible tool use and physical reasoning" in PNAS with Kevin Smith and Josh Tenenbaum. The paper presents a new domain for testing physical problem solving in humans and machines, and a model with structured object-oriented priors mimics human performance on it.
2020

Research scientist at DeepMind

She interns at Google DeepMind during her PhD, and the internship leads to a full-time job there after graduation in 2020. She works as a research scientist at DeepMind for four years, in San Francisco and London.
2021–2022

MIT PhD and Glushko Prize

She completes a PhD in Brain and Cognitive Sciences at MIT in 2021 under Josh Tenenbaum in the Computational Cognitive Science group. Her thesis, "Learning to act with objects, relations and physics", wins a Glushko Dissertation Prize from the Cognitive Science Society in 2022.
2022–2023

Learned physics simulators

In 2022, she is a first author, with Tatiana Guevara and Kimberly Stachenfeld as equal contributors, of "Physical Design using Differentiable Learned Simulators" at NeurIPS, which takes gradients through learned graph network simulators to design physical objects such as an airfoil. In 2023, her ICLR spotlight paper introduces the face interaction graph network (FIGNet), up to 2x more accurate than tuned analytic simulators such as MuJoCo and PyBullet when fit to real world data.
2025

Joining UBC and Vector

She joins UBC Computer Science in 2025 as an assistant professor, with an appointment in psychology as well. In April 2025, the Vector Institute welcomes her as its newest Faculty Member, and in July 2025, CIFAR names her a Canada CIFAR AI Chair at Vector. Her research program builds AI that learns world models through simulators trained on real-world data.
May 2026

A CIFAR Global Scholar

In May 2026, CIFAR chooses her as one of 15 Global Scholars for 2026–2028 in Learning in Machines & Brains, from over 450 applications from 41 countries. She is an assistant professor in UBC's departments of computer science and psychology.