Manipulating deformable objects such as ropes, cloth, or granular materials is hard for robots because the same object can take many shapes and the path to a target configuration is not obvious. Most prior methods predict either the robot's actions or only the robot's motion, which can miss the object's response. DefFlow instead learns to predict the goal-conditioned 3D point flow of both the robot and the deformable object from demonstrations, capturing both high-level decisions like when to grasp or release and the precise continuous motions needed to reach a goal.
The authors evaluated DefFlow across rope, granular-media, and cloth manipulation settings with diverse initial and target configurations. It substantially outperformed goal-conditioned action-prediction and robot-flow-only baselines on success rates and goal-distance metrics. The key advantage is that jointly modeling the object's flow forces the system to reason about the effect of each manipulation step, rather than just imitating motions.
DefFlow was also tested on a real robot using real point-cloud observations and goal configurations not seen during training, showing that the approach can transfer beyond simulation. The work points toward more flexible robotic manipulation in unstructured environments, though the paper does not report comparisons with other deformable-object-specific methods or discuss failure cases in detail.