Reproduce “Demonstration-guided Optimal Control for Long-term Non-prehensile Planar Manipulation” 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
Teng Xue, Hakan Girgin, Teguh Santoso Lembono, Sylvain Calinon
Researchers at Idiap Research Institute and EPFL developed a demonstration-guided optimal control framework for long-term non-prehensile planar manipulation, achieving an 85% success rate in offline planning and demonstrating robust online tracking with a Franka Emika robot across various tasks and disturbances.
Key insights
- Human demonstrations, when integrated as soft cost constraints, significantly improve the ability of optimal control algorithms to find better local optima for non-convex, hybrid contact problems, compared to direct initialization or unguided search.
- Extending the dynamics model to include separation and flexible face-switching is crucial for enabling versatile and efficient long-term non-prehensile manipulation strategies.
- A hierarchical approach that first broadly guides optimization with demonstrations and then refines the solution provides a computationally efficient alternative to complex mixed-integer programming for hybrid task and motion planning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.