Reproduce “SoftVTBench: A Deformation-Aware Visuo-Tactile Dataset and Benchmark for Deformable-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
Bowen Jing, Mingxin Wang, Ruiyang Hao, Chenchen Ge, Hanwen Shen, Junjie He, Yang Cui, Yiming Hou and 10 more
SoftVTBench introduces a visuo-tactile dataset and benchmark for deformable-object manipulation, enabling the objective evaluation of physical interaction quality beyond mere task completion. It quantifies excessive deformation using a Deformation-aware Success Rate (DSR) metric, providing a diagnostic tool for developing policies that handle delicate objects without damage.
Key insights
- A consistent discrepancy exists between task completion (TSR) and physical interaction quality (DSR), with policies frequently completing tasks while causing excessive, unsafe deformation.
- The 'cost' of handling deformable objects is not uniform across policies and task structures, and can sometimes be outweighed by other factors like target grounding or spatial generalization.
- Apparent performance gains from tactile sensing can be confounded by control granularity; a finer action space (e.g., continuous gripper control) often contributes to better interaction quality independently.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.