Reproduce “Learning to Play with Intrinsically-Motivated Self-Aware Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nick Haber, Damian Mrowca, Li Fei-Fei, Daniel L. K. Yamins
This research from Stanford University introduces an intrinsically-motivated, self-aware agent architecture that learns to play in a 3D physical environment without external rewards. By predicting and maximizing its own world-model's learning difficulty, the agent autonomously develops complex behaviors like object attention and multi-object gathering, yielding improved dynamics prediction and highly transferable visual features for tasks such as object detection and recognition.
Key insights
- An agent's 'self-awareness' of its own learning progress, modeled by predicting its world-model's loss, enables it to actively seek out challenging interactions, fostering a 'virtuous cycle' of self-driven learning.
- An adversarial relationship between a self-model and a world-model can formalize curiosity, guiding agents to 'self-curricularize' and progressively explore more complex aspects of their environment.
- Complex, human-infant-like developmental milestones, such as ego-motion mastery followed by object attention and multi-object play, can spontaneously emerge from simple intrinsic motivations in a 3D physical setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.