![[2502.18615] A Distributional Treatment of Real2Sim2Real for Object-Centric Agent Adaptation in Vision-Driven Deformable Linear Object Manipulation](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2502.18615] A Distributional Treatment of Real2Sim2Real for Object-Centric Agent Adaptation in Vision-Driven Deformable Linear Object Manipulation
Summary
This article presents a novel framework for adapting object-centric agents in manipulating deformable linear objects using visual perception and likelihood-free inference.
Why It Matters
The research addresses the challenge of transferring learned manipulation skills from simulation to real-world applications, which is crucial for advancing robotics and automation. By demonstrating effective zero-shot adaptation, this work could significantly enhance the efficiency and reliability of robotic systems in dynamic environments.
Key Takeaways
- Introduces an end-to-end framework for Real2Sim2Real adaptation.
- Utilizes likelihood-free inference to simulate the behavior of deformable linear objects.
- Demonstrates zero-shot deployment of sim-trained policies in real-world scenarios.
- Evaluates the effectiveness of domain randomization in policy learning.
- Highlights the implications of domain distributions on real-world performance.
Computer Science > Robotics arXiv:2502.18615 (cs) [Submitted on 25 Feb 2025 (v1), last revised 25 Feb 2026 (this version, v3)] Title:A Distributional Treatment of Real2Sim2Real for Object-Centric Agent Adaptation in Vision-Driven Deformable Linear Object Manipulation Authors:Georgios Kamaras, Subramanian Ramamoorthy View a PDF of the paper titled A Distributional Treatment of Real2Sim2Real for Object-Centric Agent Adaptation in Vision-Driven Deformable Linear Object Manipulation, by Georgios Kamaras and Subramanian Ramamoorthy View PDF HTML (experimental) Abstract:We present an integrated (or end-to-end) framework for the Real2Sim2Real problem of manipulating deformable linear objects (DLOs) based on visual perception. Working with a parameterised set of DLOs, we use likelihood-free inference (LFI) to compute the posterior distributions for the physical parameters using which we can approximately simulate the behaviour of each specific DLO. We use these posteriors for domain randomisation while training, in simulation, object-specific visuomotor policies (i.e. assuming only visual and proprioceptive sensory) for a DLO reaching task, using model-free reinforcement learning. We demonstrate the utility of this approach by deploying sim-trained DLO manipulation policies in the real world in a zero-shot manner, i.e. without any further fine-tuning. In this context, we evaluate the capacity of a prominent LFI method to perform fine classification over the parametric set of DLOs,...
