Reproduce “ASID: Active Exploration for System Identification in Robotic Manipulation” 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

Marius Memmel, Andrew Wagenmaker, Chuning Zhu, Patrick Yin, Dieter Fox, Abhishek Gupta

The ASID framework, developed at the University of Washington, introduces an active exploration method to autonomously refine physics simulator parameters using minimal real-world data. This enables zero-shot transfer of policies learned in the updated simulator to real-world robotic manipulation tasks, achieving success rates of 6 out of 9 for rod balancing and 7 out of 10 for shuffleboard after collecting only a single real-world exploration episode.

Key insights

  • Learning an effective exploration strategy, which is less sensitive to initial parameter inaccuracies, can be decoupled from learning a precise task-solving policy.
  • Utilizing the Fisher information matrix effectively quantifies the informativeness of trajectories for system identification, enabling targeted and efficient data collection.
  • Optimization-based system identification methods are more robust and stable than learned estimation modules, particularly when faced with limited real-world data.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.