Danijar Hafner, a prominent figure in the AI landscape, has embarked on a new venture in San Francisco’s SoMa district with a focus on developing intelligent agents capable of navigating the unpredictable. Currently operating in stealth mode, his startup is still building its identity, as evidenced by the sparse furnishings in the office, which is predominantly filled with humanoid robots. These robots, imported from China, represent the next phase in Hafner’s longstanding mission to enhance AI’s ability to adapt to uncharted environments, a crucial step for deploying robots in human settings such as homes.
To facilitate this adaptability, Hafner employs model-based reinforcement learning—a method that allows AI to create ‘world models’ that simulate physical realities. Through training within these models, agents learn to anticipate future scenarios and formulate actions, akin to dreaming about possibilities. This innovative approach significantly reduces the need for traditional trial-and-error methods in robotics, enabling agents to tackle complex tasks efficiently. Hafner’s previous successes include notable projects like Dreamer 2, which achieved human-level performance in Atari 2600 games, and Dreamer 4, which learned to mine diamonds in Minecraft using previously recorded gameplay, without direct interaction. Now, his latest initiative seeks to transition these capabilities from virtual realms to real-world applications, as he aims to solve challenges that could have a transformative impact on society.
Source: This AI entrepreneur is developing agents that can plan ahead for the unexpected via MIT Technology Review