Reproduce “A Distributional Treatment of Real2Sim2Real for Object-Centric Agent Adaptation in Vision-Driven Deformable Linear Object 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
Georgios Kamaras, Subramanian Ramamoorthy
An integrated Real2Sim2Real framework combines likelihood-free inference for accurate physical parameter estimation of deformable linear objects (DLOs) with informed domain randomization to achieve object-centric adaptation in robotic manipulation. Policies trained with this approach exhibit distinct behavioral patterns, adapting to specific DLO properties in real-world scenarios.
Key insights
- Likelihood-Free Inference combined with distributional embeddings can effectively classify subtle variations in DLO stiffness, though it shows higher uncertainty in length estimation.
- Using fine-grained posterior distributions from LFI for Domain Randomisation guides policy training towards object-specific properties, resulting in more targeted simulated environments.
- Policies trained with this approach develop distinct behavioral adaptations and object-centric control strategies, as evidenced by analysis of end-effector trajectories, even when scalar rewards are similar across policies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.