From Prediction to Action: Learning Physical Context for Deformable Linear Object Manipulation
From Prediction to Action: Learning Physical Context for Deformable Linear Object Manipulation
This research investigates the distinction between predictive utility and decision utility in manipulating deformable linear objects (DLOs). The authors introduce Geometry-Pointer-Flow (GPF), a geometry-grounded policy that separates grasp selection from motion generation. While ground-truth physical parameters improve prediction accuracy, interaction-history representations provide better control performance. Specifically, history-conditioned GPF reduces executed shape error by 9.9% relative to geometry-only GPF across simulation trials and yields the strongest zero-shot performance on hardware. All policies are trained exclusively in simulation and transferred to a physical robot without real-world training or fine-tuning. The results suggest that physical context should be evaluated based on the decisions it supports rather than predictive accuracy alone.
The study distinguishes between predictive accuracy and decision utility in manipulating deformable linear objects, demonstrating that interaction history representations yield superior zero-shot hardware performance compared to ground-truth or estimated physical parameters despite lower predictive metrics.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- Environment
- SimulationOpen source S5
Reported numbers
parameter count ratio
18.7
View original evidence
ometry-grounded policy that separates grasp-location selection from grasp-conditioned motion generation. Its geometry-only variant reduces one-step executed shape error by 32.8% relative to a DLO-specific diffusion policy, while using 18.7 times fewer parameters and achieving approximately 193 times faster inference. On this backbone, we compare physical context represented by ground-truth parameters, estimated param
Open source S3inference speed ratio
≈193
View original evidence
ometry-grounded policy that separates grasp-location selection from grasp-conditioned motion generation. Its geometry-only variant reduces one-step executed shape error by 32.8% relative to a DLO-specific diffusion policy, while using 18.7 times fewer parameters and achieving approximately 193 times faster inference. On this backbone, we compare physical context represented by ground-truth parameters, estimated param
Open source S3
- code: Reported pending Open source S6
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
ditioned GPF reduces executed error by 9.9% relative to geometry-only GPF across 470 paired simulation trials. In closed loop, interaction history improves strict-threshold attainment in simulation and yields the strongest zero-shot performance on hardware. All policies are trained exclusively in simulation and transferred to the physical robot without real-world training or fine-tuning. These results show that physi
Open source S5
ometry-grounded policy that separates grasp-location selection from grasp-conditioned motion generation. Its geometry-only variant reduces one-step executed shape error by 32.8% relative to a DLO-specific diffusion policy, while using 18.7 times fewer parameters and achieving approximately 193 times faster inference. On this backbone, we compare physical context represented by ground-truth parameters, estimated param
Open source S3
eters, and learned interaction-history representations, for both forward prediction and executed control. For prediction, ground-truth physical parameters reduce next-shape error by 15.8% over geometry alone, and estimated parameters are slightly more predictive than interaction history. This ordering does not carry over to control: estimated parameters provide no detectable one-step control gain, whereas history-con
Open source S4
cal context should be evaluated by the decisions it supports rather than by predictive accuracy alone. Code will be released upon acceptance.
Open source S6
Source：arXiv Robotics — research abstracts · arxiv.org