RoboPacker: An Autonomous Robotic Packing System for General Objects
Zhenyu Wu, Ziwei Wang, Sichao Huang, Zhan Liu, Xiuwei Xu, Haibin Yan, Jiwen Lu
AI summary
Problem
Current robotic packing systems struggle in real-world settings due to severe object occlusion, inability to generalize to novel object shapes, and inefficient planning for maximizing space utilization in confined boxes.
Approach
The system estimates complete 3D object shapes from partial views using open-vocabulary retrieval, reduces occlusion through uncertainty-driven interactive pushing, and uses hierarchical reinforcement learning to optimize packing sequences and placements.
Key results
- Open-vocabulary shape estimation for seen and novel objects
- Uncertainty-driven interactive pushing to reduce occlusion
- Hierarchical RL optimizes packing sequence and placement
- 73.3% real-world success rate packing 20 cluttered objects
Why it matters
Provides a deployable, high-throughput automation solution for e-commerce and logistics warehouses, significantly reducing manual labor costs and packing errors.
Abstract
In this paper, we propose an autonomous robot packing system named RoboPacker designed to tightly store cluttered general objects into shipping boxes with high space utilization, which is a fundamental process in numerous indus- trial applications. However, achieving tight packaging for general objects often demands significant labor from human packers, particularly in high-throughput scenes. Compared to existing robot packing approaches, RoboPacker effectively overcomes challenges such as diverse object appearances, severe occlusion, and crowded packing spaces. Specifically, we propose an open- vocabulary shape estimation method to reconstruct complete point clouds for cluttered objects. We also design effective interactions with object clutter to gather informative visual clues for shape estimation under high uncertainty. Additionally, we introduce a hierarchical reinforcement learning framework to optimize packing order, location, and orientation for maximum space utilization. The robotic packing system integrates these techniques with feasible manipulation methods for real-world im- plementation. In this way, RoboPacker achieves efficient packing of novel and irregular objects, which is more suitable for real deployment environments. The Real-world experiments demon- strate RoboPacker can tightly pack 20 densely cluttered everyday objects from 8 seen and 4 novel classes into the 40×40×20 cm shipping box with a 73.3% success rate. The demonstration video can be found at https://gary3410.github.io/RoboPacker/. Note to Practitioners—In fields such as logistics warehousing and manufacturing, which have stringent requirements for high productivity, efficiently packing disordered items into limited spaces is a core operational process. However, manually packing objects generates high labor costs due to long hours of high- intensity work. While robots handle some tasks, fully automated general item packaging remains challenging due to variable object shapes, occlusions, and space constraints. The RoboPacker industrial-grade system addresses these issues by optimizing space use and reducing human reliance, which employs ad- vanced computer vision and reinforcement learning to achieve general object packing in complex scenes. The modular design of RoboPacker integrates easily with existing robot arms, making it ideal for e-commerce, food, and pharmaceutical industries seeking intelligent upgrades. Future work will focus on expanding the system to enable more accurate and safer autonomous object packaging by incorporating additional sensors (e.g., tactile).