SoftVTBench is a new benchmark for robotic manipulation of deformable objects that prioritizes safety. It evaluates not only whether tasks are completed successfully but also whether the robot maintains physical interaction with the object without causing damage or slipping. The benchmark includes a range of scenarios and uses a combination of visual, tactile, and proprioceptive data to assess performance. Experiments show that focusing solely on task completion can lead to overestimation of policy performance, as many successful rollouts still violate safety constraints. Adding tactile sensing improves safety outcomes while maintaining comparable task success rates. --- Why it matters: This matters because robotic manipulation of deformable objects is a challenging and important problem in AI research, with applications in areas like healthcare and manufacturing. SoftVTBench provides a standardized evaluation framework that can help researchers develop safer and more reliable policies for these tasks. Source: https://arxiv.org/abs/2607.04234