WorldContact boosts robot learning speed and task success rates
WorldContact: A Contact-Centric World Model for Scalable Robot Learning
Robotics
Summary
Robots need practice to learn how to handle new objects, but getting enough practice data can be slow and expensive. The authors created WorldContact, a model that simulates how soft objects move during contact more efficiently by using bigger time steps. This lets them quickly generate more training data for robot learning. When they used this data to improve a robot’s ability to carry shopping bags, the robot became much better at succeeding on the first try.
What this means in practice
- •For robotics engineers: Speed up training data generation for robots manipulating soft or flexible objects without sacrificing dynamic accuracy.
- •For automation developers: Improve robot manipulation policies by augmenting limited simulation data with more realistic and efficiently generated contact dynamics trajectories.
Authors
Caoliwen Wang, Mengdi Wang, Heng Zhang, Shixun Huang, Siyuan Chen, Chao Liu, Anpei Chen, Zhendong Wang, Peter Yichen Chen, Huamin Wang
Abstract
Adapting robots to new objects and tasks requires interaction experience that can be costly to obtain. We present WorldContact, a contact-centric world model for deformable-object manipulation, constructed from a limited set of high-quality trajectories to generate additional training data efficiently. It predicts object dynamics using larger time steps than the source numerical simulator, which requires small integration steps to resolve rapid motion and prevent interpenetration. We evaluate WorldContact across 16 shopping-bag manipulation tasks. State-rollout measurements on a single H100 GPU show a $10\times$ speedup over the source simulator, excluding rendering and disk I/O. We use the generated data to fine-tune an existing vision-language-action policy and deploy it directly on a real robot. In bag lifting, the same policy achieves 65% single-attempt success when fine-tuned on source simulation data alone, compared with 95% when fine-tuned on the dataset expanded with WorldContact. These results support efficient data generation with WorldContact for robot policy adaptation.