Reproduce “Distributionally Robust Control via Stein Variational Inference for Contact-Rich 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
Hrishikesh Sathyanarayan, Victor Vantilborgh, Harish Ravichandar, Tom Lefebvre, Ian Abraham
A new control framework called Stein Variational Distributionally Robust Optimization (SV-DRO) integrates Stein Variational Inference with model-based control to adaptively model task-sensitive parameter uncertainty in contact-rich robotic manipulation. The approach achieves up to a 3x improvement in robustness and higher task success rates (e.g., 93.25% for bimanual pushing and 100% for within-hand positioning) compared to existing methods, even with broad uncertainty and limited feedback.
Key insights
- Adaptively shaping the uncertainty distribution through Stein Variational Inference (SVI) based on task-aware sensitivities allows the controller to prioritize parameter variations most critical for task success.
- The deterministic evolution of parameter particles via SVI provides more stable and consistent convergence for Distributionally Robust Optimization (DRO) objectives compared to stochastic sampling.
- Robust manipulation can be achieved by modeling task-relevant uncertainty through an optimality gap, enabling effective control even with broad initial parameter uncertainty and limited feedback.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.