Synthetic Data for Soft Body Manipulation

thinkingsquishy: exploring synthetic data generation techniques for soft-body manipulation tasks.


Exploring possible synthetic data generation techniques based off of Dittogym for soft body manipulation tasks. See the code here. Currently a work in progress.

Generalist robots generally struggle with precise deformable object manipulation due to a variety of reasons, including hardware limitations (e.g. gripper design), sensory inadequacies (most models omit tactile feedback in favor of image input), but also a lack of high-quality in-distribution data. This work leverages existing physics-accurate soft robot policies, enforcing policy movement restrictions in order to generate ideal trajectories completable by prehensile manipulation.