Reproduce “Convolutional Occupancy Models for Dense Packing of Complex, Novel 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

Nikhil Mishra, Pieter Abbeel, Xi Chen, Maximilian Sieb

F-CON, a frustum-convolutional occupancy network, combined with the COB-3D-v2 dataset, enables real-world robotic systems to densely pack complex, novel objects from cluttered bins. The system achieves up to an 88% success rate and a packed volume of 0.440 in real-world experiments, outperforming prior shape completion methods.

Key insights

  • Accurate 3D shape completion from partial observations is a fundamental perceptual bottleneck, and addressing it directly can significantly improve real-world dense packing performance.
  • A frustum-convolutional architecture, which processes camera-centric and object-centric feature volumes, effectively learns strong geometric priors for robust generalization to novel objects and cluttered scenes.
  • Focusing on superior perception enables simpler, existing planning methods to achieve enhanced performance in complex robotic manipulation tasks, bridging the sim-to-real gap.

Summary by alphaXiv.

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