Reproduce “Learning Awareness Models” 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
Brandon Amos, Laurent Dinh, Serkan Cabi, Thomas Rothörl, Sergio Gómez Colmenarejo, Alistair Muldal, Tom Erez, Yuval Tassa and 2 more
Researchers from Carnegie Mellon University, University of Montreal, and DeepMind developed a method for agents to learn persistent representations of external objects by exclusively predicting their own proprioceptive and tactile sensor feedback. This approach, which includes active exploration strategies, enabled model-based control to achieve tasks requiring awareness of unobserved object properties in both simulated and real robotic hands.
Key insights
- An agent can form 'holistic, persistent representations' of external objects in its environment by accurately predicting its own internal body sensations, even without direct external observation or explicit labels.
- Actively exploring by maximizing predictive uncertainty allows an agent to gather more informative data, leading to superior dynamics models that are better at anticipating and reasoning about external objects.
- These implicitly learned awareness models can be effectively leveraged by Model Predictive Control for tasks requiring interaction with unobserved object properties, demonstrating their utility for autonomous robot control.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.