Reproduce “Monte-Carlo Tree Search for Efficient Visually Guided Rearrangement Planning” 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
Yann Labbé, Sergey Zagoruyko, Igor Kalevatykh, Ivan Laptev, Justin Carpentier, Mathieu Aubry, Josef Sivic
An integrated system combines Monte-Carlo Tree Search with a calibration-free multi-object visual state estimation system. It achieves efficient and scalable visually guided rearrangement planning for numerous objects using an uncalibrated RGB camera and synthetic training data.
Key insights
- MCTS effectively balances exploration and exploitation in rearrangement planning, enabling it to scale to a higher number of objects and find plans with fewer moves than heuristic methods.
- Predicting object positions directly in the robot's coordinate frame, with the robot visible in the image, allows for robust, calibration-free visual state estimation from an uncalibrated camera.
- Domain randomization applied to synthetic data is sufficient for training a generalizable multi-object visual perception system, reducing reliance on real-world data collection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.