SIM-TO-REAL · GENTLE MANIPULATION
Cross-Category Deformable and Fragile Object Grasping from Stress-Aware Demonstrations at Scale
Learning to grasp fragile and deformable objects from stress-aware demonstrations generated at scale in simulation.
StressLess uses finite-element stress to synthesize stable, gentle grasps at scale, then trains one point-cloud policy that transfers zero-shot across fragile and deformable objects and improves both lifting success and grasp gentleness (object integrity) in real-world trials.
01 / PROJECT VIDEO
Highlights · motivation · method · experiments
02 / METHOD
A finite-element contact model scores grasp pose and closing width for internal stress and grip stability. Selected grasps run in soft-body simulation across randomized object poses, sizes, shapes, and materials.
One Diffusion Policy learns action chunks from all simulated demonstrations. A PointNet encoder processes the depth point cloud; robot pose and gripper width provide proprioception.
The policy controls a 7-DoF arm and parallel-jaw gripper from one fixed depth camera. It predicts end-effector pose and gripper width without tactile or force feedback.
03 / QUALITATIVE EVALUATION
When an initial grasp fails, StressLess can reopen, reapproach, and grasp again. The shown RGB-frozen baseline stops after its first failed attempt.
In 31 first-attempt failures, StressLess retried 25 times and recovered 10 lifts.
In this mushroom example, the RGB-scratch baseline damages the object during grasping; StressLess lifts it while keeping it intact.
The paper measures gentleness with a lift-conditioned safety score assigned by two blinded evaluators.
The simulated set contains standing cylinders but no lying can. StressLess handles the real can in this example, with a retry; the RGB-scratch baseline damages it.
Across can trials, StressLess achieved Lift 1.00 and Safe 0.88; the paper also evaluates a thin-shell paper cup.
Real-world evaluation objects across 11 categories, including the can and paper cup.
FULL PAPER
StressLess: Cross-Category Deformable and Fragile Object Grasping from Stress-Aware Demonstrations at Scale