Latent Diffeomorphic Co-Design of End-Effectors for Deformable and Fragile Object Manipulation

Venue: RSS 2026 (Sydney, Jul 13–17) · Session: Robot & Sensor Design · paper #195 Authors: Kei Ikemura, Yifei Dong, Florian T. Pokorny arXiv: 2602.17921 · program page

Summary compiled from the arXiv paper (v1); all numbers quoted from the paper. Trend context: RSS 2026 survey.

Pipeline: (a) a diffeomorphic design space parameterizes physically valid end-effector geometries; (b) a CMA-ES evolutionary optimizer searches designs d to maximize a stress-aware score; (c) for each design a design-conditioned controller is optimized from privileged simulator signals; (d) rollouts of the privileged policy are distilled into a 3D diffusion policy on pointclouds; (e) the distilled policy is deployed zero-shot on the real robot.*

Problem

Manipulating deformable and fragile objects (jelly, raw fish fillet) demands both the right gripper geometry and gentle reactive control, but prior work optimizes hardware or control in isolation. The authors present the first co-design framework that jointly optimizes end-effector morphology and control for deformable/fragile object manipulation (DFOM).

Method

Three contributions: (1) Latent DiffeoMorphism (LDM) — a shape parameterization that represents end-effectors as smooth invertible deformations (stationary velocity field flow) of a base mesh, giving an expressive but low-dimensional (~15-D) design space that avoids self-penetration; (2) a stress-aware bi-level co-design pipeline — an outer CMA-ES loop searches designs to maximize a score trading task success against soft-body stress (mean and top-percentile σ_max from the simulator), with an inner design-conditioned motion-primitive controller using privileged stress/centroid signals to trigger phase transitions; (3) privileged-to-pointcloud distillation — 100 expert trajectories per task train a student 3D diffusion policy on segmented pointclouds for zero-shot real deployment under domain randomization. Simulation is Genesis with MPM soft-body physics; baselines are parallel-jaw (PJ), Bayesian-optimization co-design (BO), and end-to-end RL co-design (RLPD).

Results

In simulation (averaged over 3 seeds), the co-designed gripper wins across jelly grasping, jelly pushing, and fillet scooping. On the oversized jelly cube it achieves 33% higher success and 36% lower induced stress than PJ; jelly cylinder grasping reaches 0.97 success (vs PJ 0.63, BO 0.80, RL 0.04); jelly pushing beats BO by 6.4% success and 6.6% stress on average. RL co-design collapses (0.04-0.28 success, very high stress) due to reward hacking. An LDM-vs-shape-primitive ablation confirms LDM's expressiveness. Real-world on a UFactory xArm 7 with an L515 LiDAR and 3D-printed designs: cylindrical jelly grasping 9/9 success (BO 5/9) and pushing 10/12 (BO 9/12), with minimal visible breakage.

Significance

Brings morphology-control co-design — long studied for locomotion — to fragile deformable manipulation, showing that "intelligent morphology can reduce control complexity"; a hardware-design counterpart to the tactile/gentle-manipulation work in Review-Tactile-VLA and Review-Dexterous-Manipulation.

← Back to RSS 2026 survey · RSS-2026-Papers · Home