Hi, I'm Danijar! My goal is to develop autonomous artificial intelligence that makes decisions given only limited human instruction.
I'm a PhD candidate at the University of Toronto with Jimmy Ba, a visiting student at UC Berkeley with Pieter Abbeel, and an intern at DeepMind. Previously, I interned at Google Brain for 6 years. I completed my MRes at UCL and the Gatsby Unit with Tim Lillicrap and Karl Friston. My work is supported by Canada's Vanier Scholarship.
Contact: mail AT danijar.com
Research
My research aims at discovering the computational principles of human thinking and replicating them to build useful artificial intelligence. I approach this challenge by designing unsupervised algorithms that learn through interaction with complex environments, going beyond task-specific agents that are inherently narrow in their abilities.
The key questions I’m currently focusing on are:
- World Models Learning large models from video to develop a general
understanding of the world and enable planning by imagining future outcomes
of potential actions.
Key papers: DayDreamer, Dreamer, DreamerV2, PlaNet - Agent Objectives Now that we understand the space of agent
objectives, we need to find scalable algorithms for exploration, empowerment,
and skill discovery.
Key papers: LEXA, Plan2Explore, APD - Temporal Abstraction While humans plan at an abstract level, control
algorithms are still limited by exploring, planing, and assigning credit at
the level of primitive actions.
Key papers: Director, LSP, ClockworkVAE
This talk gives a good technical overview of my research.
Featured Media
See YouTube for more videos.
-
Learning to Walk in the Real World in 1 Hour
-
Unsupervised Intelligent Agents Technical Talk
-
Interview in Daily Mail
-
Interview in MIT Technology Review
-
Lab Spotlight in Tech Crunch
-
LEXA Paper Overview
-
Open-Endedness Panel
-
Talk RL Podcast
-
Introducing Director
-
Introducing DreamerV2
-
Introducing Dreamer
-
Introducing Plan2Explore
Highlighted Work
See Google Scholar for more publications.
-
Learning to Model the World with Language
ICML 2024 (oral)
-
Evaluating Long-Term Memory in 3D Mazes
ICLR 2023
-
Deep Hierarchical Planning from Pixels
NeurIPS 2022
-
Discovering and Achieving Goals via World Models
NeurIPS 2021 (26%), URL 2021 (oral), SSL 2021 (oral)
-
Benchmarking the Spectrum of Agent Capabilities
ICLR 2022, DRLW 2021 (oral)
-
Clockwork Variational Autoencoders
NeurIPS 2021 (26%)
-
Latent Skill Planning for Exploration and Transfer
ICLR 2021 (28%)
-
Mastering Atari with Discrete World Models
ICLR 2021 (28%)
-
Planning to Explore via Self-Supervised World Models
ICML 2020 (22%)
-
Dream to Control: Learning Behaviors by Latent Imagination
ICLR 2020 (oral, 4%), DRLW 2019 (oral)
-
A Deep Learning Framework for Neuroscience
Nature Neuroscience
-
Noise Contrastive Priors for Functional Uncertainty
UAI 2019 (26%)
-
Learning Latent Dynamics for Planning from Pixels
ICML 2019 (23%)
-
Sample-Efficient Reinforcement Learning with Stochastic Ensemble Value Expansion
NeurIPS 2018 (oral, 0.6%)
-
Sim-to-Real: Learning Agile Locomotion For Quadruped Robots
RSS 2018 (31%)
Short Biography
Danijar Hafner is a PhD candidate in artificial intelligence at the University of Toronto with Jimmy Ba and a visiting student at the University of California, Berkeley with Pieter Abbeel. He is also a research scientist intern at DeepMind. Danijar’s research aims at building intelligent machines based on the computational principles of the brain. Towards this goal, he focuses on scaling up embodied artificial intelligence through general world models, unsupervised objectives, and deep reinforcement learning. Danijar completed his MRes in Computational Statistics and Machine Learning at UCL with Tim Lillicrap and Karl Friston. His work is supported by Canada’s Vanier Scholarship.