Yunzhu Li (@YunzhuLiYZ)
2024-07-11 | โค๏ธ 170 | ๐ 31
Check out our RSS2024 paper (also the Best Paper Award at the ICRA2024 deformable object manipulation workshop) on dynamics modeling of diverse materials for robotic manipulation. ๐ค
We considered a diverse set of objects, including ropes, clothes, granular media, and rigid objects with different mass distributions.
The key to this capability is again the Graph-Based Neural Dynamics (GBND), which extends our series of work on DPI-Net, RoboCraft, RoboCook, and RoboPack. We conditioned GBND with material parameters and used it for inverse problems such as (1) material adaptation and (2) model-based planning.
Key Takeaways:
- Graph-based structured scene representations show surprising effectiveness in capturing the scene dynamics for diverse materials.
- Material conditioning allows precise modeling and effective manipulation of objects in the same category but with distinct dynamics (e.g., stiff ropes like cables and soft ropes like yarn and shoelaces).
- Huge potential still lies ahead in extending to even larger-scale and heterogeneous environments.
Kudos to @kaiwynd and @BaoyuLi6 for leading this project. Check out Kaifengโs thread for more details and come chat with us at RSS2024!
๋ฏธ๋์ด
์ธ์ฉ ํธ์
Kaifeng Zhang (@kaiwynd)
(1/8) Introducing AdaptiGraph, our new paper accepted by RSS2024
We show that a GNN dynamics model can model deformable objects with varying physical properties.
Website: https://t.co/Fpb09qA2bE ArXiv: https://t.co/2xehBgU2gj
Work done w/ @BaoyuLi6, Kris Hauser, @YunzhuLiYZ https://t.co/RqZOioyr1Q
๐ฌ ์์
