Demonstration augmentation is proposed for cost-efficient data acquisition, but existing methods are fundamentally limited in deformable manipulation due to two challenges: (1) the state space is high-dimensional with physics-induced constraints, making valid configurations impossible to reach via low-dimensional pose perturbations; and (2) trajectory transfer is non-equivariant, as material points no longer move rigidly together under deformation.
We present DeformGen, a dynamics-based augmentation framework that achieves topological diversity for deformable objects. For the state challenge, DeformGen expands the valid state distribution by applying localized physical disturbances and forward-simulating the dynamics to obtain topology-coherent, physically plausible deformable states.
For the trajectory challenge, DeformGen transfers source manipulation trajectories via deformation-field warping, which lifts per-particle displacements into a continuous spatial function to adapt the end-effector trajectory consistently with the deformed geometry.
In this way, our method jointly augments the state distribution and its associated manipulation behavior. Experiments on high-fidelity deformable manipulation benchmarks show that DeformGen generally improves policy learning compared with training on the original demonstrations alone and with rigid-style augmentation baselines.
DeformGen first expands the physically plausible state distribution, then synthesizes manipulation trajectories that remain consistent with each deformed geometry.
The objective is to generate diverse object configurations for the same task while staying inside the physically plausible subspace of the deformable object. Rather than perturbing a low-dimensional object pose or applying arbitrary per-particle noise, DeformGen starts from a valid state, applies localized physical disturbances, and forward-simulates the calibrated dynamics so the simulator continuously enforces material and contact constraints.
The objective is to synthesize valid manipulation trajectories for unseen object configurations. DeformGen constructs a closed-form deformation field from per-particle displacements between source and deformed states. For each waypoint, K-nearest-neighbor particle displacements are interpolated to obtain the position offset, while a local Jacobian estimates the orientation update. The resulting approach, grasp, and manipulation phases are replayed to retain only executable trajectories.
Success rate (%) on held-out deformable-object states. DG denotes the full DeformGen pipeline.
DG average success
DG average success
DG average success
DG average success
| Policy | 1 Source | SMG* | DG* | DG |
|---|---|---|---|---|
| ACT | 1.33 | 48.17 | 46.83 | 59.00 |
| Diffusion Policy | 2.33 | 38.00 | 41.00 | 37.33 |
| SmolVLA | 2.50 | 40.33 | 55.00 | 56.50 |
| π0 | 2.33 | 27.83 | 51.67 | 56.67 |
Generated supervision is validated through simulation replay before policy training.
@misc{lin2026deformgen,
title={DeformGen: Dynamics-Based Topology Augmentation for Deformable Manipulation Policy Learning},
author={Zili Lin and Wenyao Zhang and Yuyang Zhang and Zekun Qi and Junyan Lin and Hanxin Zhu and Jiaolong Yang and Zhibo Chen and Yao Mu and Xiaokang Yang and Xin Jin and Wenjun Zeng},
year={2026},
eprint={2606.25939},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2606.25939},
}