Danijar Hafner builds AI agents that plan ahead — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Danijar Hafner founded a stealth AI startup in San Francisco in fall 2025, aiming to develop agents that can navigate novel physical environments using advanced world models.
- Danijar Hafner employs model-based reinforcement learning to train AI agents, enabling them to simulate and predict outcomes in unseen scenarios—effectively allowing robots to 'dream' or imagine future actions before executing them.
- Danijar Hafner demonstrated the power of his approach through a series of AI agents—PlaNet, Dreamer 2, Dreamer 3, and Dreamer 4—that achieved human-level performance in Atari games and autonomously mined diamonds in Minecraft using only offline data.
- Danijar Hafner transitioned his work from virtual simulations to physical robotics with the DayDreamer project, where robots used Dreamer algorithms to react to real-world disruptions like being pushed over without specific retraining.
- Google DeepMind colleague Timothy Lillicrap described Hafner as being in the top half of 1% of researchers he’s worked with, noting that Hafner often built complex systems single-handedly that would typically require entire engineering teams.
Why it matters: Hafner’s approach reduces the need for costly real-world trial-and-error training in robotics, making autonomous adaptation faster and more scalable. This could accelerate deployment of general-purpose robots in unstructured human environments like homes or hospitals, where flexibility is critical and pre-programming every scenario is impossible.
Ask SkimNews
