Robotic bin packing aids in a wide range of real-world scenarios, such as e-commerce and warehouses. Yet, existing works focus mainly on considering object shapes to optimize packing compactness and neglect object properties such as fragility, edibility, and chemistry that humans typically consider when packing objects. This work proposes OPA-Pack, the very first robot packing framework that accounts for object properties. OPA-Pack advances robotic bin packing by avoiding closely packing incompatible object pairs and reducing pressure on fragile objects while optimizing compactness. This work is conditionally accepted to Transactions on Robotics (T-RO).
OPA-Pack Object-Property-Aware Robotic Bin Packing
Researchers
Introduction
The Main Impact
1
OPA-Pack is a new paradigm for object-property-aware robotic bin packing. It is designed to avoid closely packing incompatible object pairs and reduce pressure on fragile objects while optimizing compactness. Technically, we introduce an object-property recognition scheme that combines retrieval-augmented generation with chain-of-thought reasoning.
OPA-Pack has two stages: (a) object-property recognition and (b) property-aware packing learning. First, in the object-property recognition, a Vision–Language Model (VLM) builds the OPA dataset using CLIP-based Retrieval-Augmented Generation (RAG) and Chain-of-Thought reasoning (CoT) to predict 11 properties and the avoidance relations for incompatible object pairs. Second, in the property-aware packing learning, OPA-Net uses these properties to reduce pressure on fragile items and separate incompatible pairs while preserving packing compactness.
Visualization of packing results for three subsets—Hardware & Stationery, Daily Necessities, and Kitchen—comparing the baseline and our method. Columns show: overall packing, fragile-object visualization, and an avoidance-pair visualization. Our method places fragile items on top to avoid compression and better separates incompatible pairs.
2
Also, we propose OPA-Net, a model designed to simultaneously separate incompatible object pairs, reduce pressure on fragile items, and maximize packing compactness. Experimental results manifest that OPA-Pack greatly improves the accuracy of separating incompatible object pairs (from 52% to 95%) and largely reduces pressure on fragile objects (by 29.4%), while maintaining good packing compactness.
3
In addition, we evaluate OPA-Pack on a real packing platform with NACHI MZ07 robot arm using 70 sequences of real objects selected from 100 object types.
Visual comparison of our method and the baseline in real-world packing results. Left: our method better prevents heavy objects from being placed on fragile persimmon and plum, addressing a limitation of the baseline. By employing our approach, we can reduce fruit damage. Right: the baseline packs the persimmon directly on top of the medical kit or next to the cockroach poison, which could contaminate the unsealed fresh fruit. In comparison, our method avoids packing them closely.