Reproduce “Physics-Driven Data Generation for Contact-Rich Manipulation via Trajectory Optimization” 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
Lujie Yang, H.J. Terry Suh, Tong Zhao, Bernhard Paus Graesdal, Tarik Kelestemur, Jiuguang Wang, Tao Pang, Russ Tedrake
The authors present a method that uses human demonstrations to guide physics-based trajectory optimization for contact-rich robot manipulation. They retarget demonstrated motions to a robot, then optimize actions under simulated dynamics and varied physical conditions. In a zero-shot iiwa box-flipping test, the generated-data policy succeeded in 17 of 23 trials, versus 6 of 23 for the original-demonstration policy; collision-model errors still caused failures.
Key insights
- Dynamically feasible trajectories, ensured by physics-driven trajectory optimization, are essential for successful execution of contact-rich manipulation tasks, as purely kinematic approaches often fail.
- Human demonstrations provide crucial global guidance, which, when combined with the local refinement capabilities of model-based trajectory optimization, overcomes limitations of each method in isolation.
- Task knowledge can be effectively transferred across vastly different robot embodiments and from human hands via VR demonstrations, reducing the need for hardware-specific data collection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.