AI xPress

AI and computer science research.

Back to Foragentis

Robotics & Embodied AI

A single reinforcement-learning algorithm outperformed specialized methods across more than 150 control tasks, a 2025 study reports

Robotics & Embodied AI November 10, 2026

A robotic prosthetic hand on a workbench

In a 2025 paper in Nature, researchers at the University of Toronto report that one algorithm, Dreamer, outperformed specialized methods across more than 150 different control tasks while using a single configuration. Dreamer learns a model of its environment and improves its behavior by imagining future scenarios. The team also reports that, to their knowledge, Dreamer is the first algorithm to collect diamonds in Minecraft from scratch without human data or curricula.

What did the researchers find about generalization?

The paper describes a reinforcement-learning algorithm that worked across more than 150 control tasks with one configuration. That means one general algorithm handled many tasks instead of needing a separate setup for each one. The authors report that it outperformed specialized methods on those tasks.

How does Dreamer learn?

Dreamer learns a model of the environment and uses that model to imagine future situations. It then improves its behavior from those imagined experiences. The paper describes normalization, balancing, and transformations as methods that keep learning stable across different domains.

What does the Minecraft result show?

The authors write that Dreamer, applied out of the box, collected diamonds in Minecraft from scratch without human data or curricula. The abstract says this required exploring farsighted strategies from pixels and sparse rewards in an open world. To the authors' knowledge, this is the first algorithm to do so.

Frequently asked questions

What is reinforcement learning?

Reinforcement learning is a training approach in which an algorithm learns by trying actions and receiving rewards. In this study, the algorithm had to solve control tasks without a new setup for each task.

Does Dreamer need separate tuning for each task?

The paper reports results with a single configuration across more than 150 tasks. The authors used the same settings across the tested tasks.

Does Dreamer need human demonstrations?

It did not need human demonstrations. The abstract reports that the Minecraft result came from scratch without human data or curricula.

Limits and what's next

The abstract describes one algorithm, the third version of Dreamer, evaluated on more than 150 control tasks and on an open-world Minecraft task. The results come from those tasks. This is one study, and the abstract reports no independent replication. The authors state the Minecraft diamond result was the first to their knowledge.

Source: Danijar Hafner et al., “Mastering diverse control tasks through world models,” Nature, 2025.