Reproduce “Unsupervised Curricula for Visual Meta-Reinforcement Learning” 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
Allan Jabri, Kyle Hsu, Ben Eysenbach, Abhishek Gupta, Sergey Levine, Chelsea Finn
The CARML framework proposes an unsupervised method for visual meta-reinforcement learning, generating adaptive task curricula directly from pixel observations. This enables agents to acquire generalizable adaptation strategies without manual task specification, achieving improved transferability and sample efficiency for downstream tasks compared to existing baselines.
Key insights
- Information maximization between state trajectories and a latent task variable (I(τ; z)) provides an intrinsic objective that balances task diversity and structure for effective unsupervised curriculum generation.
- A hybrid discriminative-generative clustering approach for modeling the task distribution `q_φ` is crucial for handling high-dimensional visual data robustly and preventing mode-collapse.
- An adaptive, co-evolving curriculum, where task discovery and meta-learning iteratively refine each other, generates more effective and diverse pre-training tasks tailored to the learner's improving capabilities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.