SoftVTBench: A Deformation-Aware Visuo-Tactile Dataset and Benchmark for Deformable-Object Manipulation
The benchmark counts manipulation as failure if the object is squeezed or slipped too far, even when the task is completed.
SoftVTBench pairs 4,000 expert demonstrations with multi-view vision, dual-finger tactile readings, proprioception, language, actions, and evaluator-only FEM deformation states. Its Deformation-aware Success Rate requires both task completion and peak normalized deformation within an object-specific tolerance. In tests across Diffusion Policy, pi_0.5, and FastWAM, some successful rollouts in every in-distribution setup still violated deformation limits. Under distribution shift, visuo-tactile variants improved task success in all six policy-suite comparisons and improved DSR in five, but the paper says touch alone did not guarantee effective fusion. HF Daily Papers' note
SoftVTBench pairs 4,000 expert demonstrations with multi-view vision, dual-finger tactile readings, proprioception, language, actions, and evaluator-only FEM deformation states. Its Deformation-aware Success Rate requires both task completion and peak normalized deformation within an object-specific tolerance. In tests across Diffusion Policy, pi_0.5, and FastWAM, some successful rollouts in every in-distribution setup still violated deformation limits. Under distribution shift, visuo-tactile variants improved task success in all six policy-suite comparisons and improved DSR in five, but the paper says touch alone did not guarantee effective fusion. HF Daily Papers' note
score 5