- Adversarial Object Rearrangement in Constrained Environments with Heterogeneous Graph Neural Networks(arXiv)
Authors: Xibai Lou, Houjian Yu, Ross Worobel, Yang Yang, Changhyun Choi
Summary: Adversarial object rearrangement in the actual world (e.g., beforehand unseen or outsized objects in kitchens and shops) may gain advantage from understanding job scenes, which inherently entail heterogeneous parts corresponding to present objects, aim objects, and environmental constraints. The semantic relationships amongst these parts are distinct from one another and essential for multi-skilled robots to carry out effectively in on a regular basis eventualities. We suggest a hierarchical robotic manipulation system that learns the underlying relationships and maximizes the collaborative energy of its various expertise (e.g., pick-place, push) for rearranging adversarial objects in constrained environments. The high-level coordinator employs a heterogeneous graph neural community (HetGNN), which causes concerning the present objects, aim objects, and environmental constraints; the low-level 3D Convolutional Neural Community-based actors execute the motion primitives. Our strategy is skilled solely in simulation, and achieved a mean success fee of 87.88% and a planning price of 12.82 in real-world experiments, surpassing all baseline strategies. Supplementary materials is obtainable at https://sites.google.com/umn.edu/versatile-rearrangement.
