Reproduce “Reinforcement Learning for Pivoting Task” 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

Rika Antonova, Silvia Cruciani, Christian Smith, Danica Kragic

A deep reinforcement learning methodology was developed at KTH Royal Institute of Technology for training robots to perform in-hand pivoting tasks with parallel grippers. Policies learned in a simple, noisy simulator successfully transferred to a real Baxter robot, achieving 93% task completion for known objects and 83% for objects with unknown properties, demonstrating robustness to simulation-to-real discrepancies.

Key insights

  • A simple simulator, by injecting noise and delays, can effectively enable the transfer of learned reinforcement learning policies to real robots, mitigating the simulation-to-real gap without requiring high-fidelity modeling.
  • Policies trained with simulated environmental noise and parameter variations exhibit strong generalization capabilities, successfully manipulating objects with substantially different and unknown physical properties on a real robot.
  • Model-free algorithms like TRPO can learn robust continuous control policies for complex manipulation tasks, such as pivoting, overcoming challenges faced by other algorithms like DDPG in initial experiments.

Summary by alphaXiv.

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