Self-Supervised Unseen Object Instance Segmentation via Long-Term Robot Interaction

TL;DR

This paper presents a robotic system that improves unseen object segmentation by leveraging long-term interactions and self-supervised learning, significantly enhancing segmentation accuracy and grasping performance in real-world scenarios.

Contribution

The system defers segmentation decisions over multiple actions and uses multi-object tracking and video segmentation for self-supervised learning, a novel approach in robotic perception.

Findings

Improved segmentation accuracy after fine-tuning with real-world data.

Enhanced grasping success on unseen objects after system training.

Effective cross-domain generalization of segmentation networks.

Abstract

We introduce a novel robotic system for improving unseen object instance segmentation in the real world by leveraging long-term robot interaction with objects. Previous approaches either grasp or push an object and then obtain the segmentation mask of the grasped or pushed object after one action. Instead, our system defers the decision on segmenting objects after a sequence of robot pushing actions. By applying multi-object tracking and video object segmentation on the images collected via robot pushing, our system can generate segmentation masks of all the objects in these images in a self-supervised way. These include images where objects are very close to each other, and segmentation errors usually occur on these images for existing object segmentation networks. We demonstrate the usefulness of our system by fine-tuning segmentation networks trained on synthetic data with real-world data collected by our system. We show that, after fine-tuning, the segmentation accuracy of the networks is significantly improved both in the same domain and across different domains. In addition, we verify that the fine-tuned networks improve top-down robotic grasping of unseen objects in the real world.