Reproduce “Pose-Agnostic Robotic Functional Grasping via Observation-Action Canonicalization” 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
Le Qiu, Cole Harrison, Jiankai Sun, Yao Liu, Suning Huang, Qianzhong Chen, Yang You, Marco Pavone
Researchers from Stanford University and Tsinghua University developed AnyMug, a framework that enables a single closed-loop policy to perform functional handle grasping for mugs, generalizing across diverse object geometries and poses. This approach, using observation-action canonicalization, achieved an 80% success rate on a real Franka Panda robot with unseen mugs, surpassing baselines like Motion Planning and Diffusion Policy.
Key insights
- Canonicalizing both observations and actions simplifies the learning problem, allowing a single policy to generalize across diverse object poses by operating in a consistent, object-centric frame.
- Explicitly rewarding finger opposition is crucial for guiding the gripper to precisely wrap around thin structures like handles, enabling functional contact and preventing empty grasps.
- A combination of pose curriculum learning, domain randomization (including mug geometry and simulated pose estimation noise), and physical parameters is essential for robust sim-to-real transfer.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.