Reproduce “NeRP: Neural Rearrangement Planning for Unknown Objects” 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
Ahmed H. Qureshi, Arsalan Mousavian, Chris Paxton, Michael C. Yip, Dieter Fox
NeRP, a deep learning-based system, performs multi-step object rearrangement planning for previously unknown objects using only perceptual data, trained entirely in simulation. It achieved a 94.56% success rate on synthetic unseen object tasks, outperforming model-based and classical baselines, and successfully generalized to real-world scenarios with varying object counts.
Key insights
- An object-centric graph representation, combined with neural networks, effectively enables robots to reason about relationships and plan multi-step actions for unknown objects using only perceptual data.
- Stochasticity in predicting object placement, particularly through dropout layers, is crucial for exploring varied placement options and resolving blocked target scenarios efficiently.
- Dedicated neural components for object selection and goal satisfaction are fundamental for producing efficient and accurate long-horizon plans, preventing suboptimal or non-terminating action sequences.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.