Reproduce “DICArt: Advancing Category-level Articulated Object Pose Estimation in Discrete State-Spaces” 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

Li Zhang, Mingyu Mei, Ailing Wang, Xianhui Meng, Yan Zhong, Xinyuan Song, Liu Liu, Rujing Wang and 2 more

DICArt formulates category-level articulated object pose estimation as discrete conditional generation rather than direct continuous regression. By combining quantized SE(3) pose tokens, an adaptive discrete-diffusion denoising process, and hierarchical joint constraints, it reduces reported rotation and translation errors across synthetic, semi-synthetic, and real-world benchmarks, including 1.7° rotation error on drawers and 8.2° average rotation error on a seven-part robot arm.

Key insights

  • Coordinated token-level denoising addresses the problem that the components of a single rotation may otherwise converge at different rates; removing the reformulated denoising process raises drawer rotation error from 1.7° to 4.0°.
  • Kinematic coupling provides physically structured predictions and helps maintain accuracy under occlusion, with drawer rotation error staying around 1.8–1.9° across visibility ranges.
  • The discrete formulation outperforms continuous diffusion in the reported ablation, reducing drawer translation error from 0.143 m to 0.072 m and rotation error from 3.1° to 1.7°.

Summary by alphaXiv.

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