Amazon Robotics Open-Sources ARMBench: A Large Open-Source Dataset For Training Robots

Picking, sorting, and packaging are just some of the many warehouse operations that may be automated using robotic object-handling systems. It is not easy to construct robust and scalable robotic systems for use in handling objects in warehouses. Warehouses now handle millions of items that vary greatly in size, shape, material, and other physical characteristics. The disorganized arrangements of these things within containers provide difficulties for robotic perception and planning. More fundamental work is required for visual perception algorithms like object segmentation and identification to work with novel items and setups. New challenges (like defect identification) and metrics (like assessing uncertainty in prediction) need to be established to capture the size and high-precision needs of such systems.

A large-scale benchmark dataset for perception and manipulation problems in robotic pick-and-place tasks is available through ARMBench, the Amazon Robotic Manipulation Benchmark. The dataset contains many goods and setups gathered in an Amazon warehouse. It includes high-quality annotated photographs and videos for the many steps of robotic manipulation, such as picking, transferring, and positioning.

Datasets ARMBench presents:

1. A trove of sensor readings gathered by a pick-and-place robotic manipulation workcell

2. Containerized object metadata and visual references

3. Annotations are collected automatically (due to the system’s architecture) or manually (by humans).

4. Robotic manipulation perception benchmarking tasks and metrics

Amazon product categories and physical attributes like size, shape, material, deformability, look, fragility, etc., are both represented in the dataset.

The data gathering platform is a warehouse pick-and-place workcell utilizing robotic manipulation. The robotic arm in the workcell is equipped with a vacuum end-effector. It shows various items with varying properties and arrangements inside a box. The robotic arm’s job is to singulize the container’s contents, taking out one item at a time and placing it on revolving trays.

When a container is empty, the workcell is released, and a new container is brought in to take its place. While the process is fully automated, a human is still involved to check in on the progress of each pick-and-place task, provide any necessary annotations, and fix any problems that arise. Many image sensors are installed in the workcell to facilitate and confirm the pick-and-place operation.

The dataset has annotated data for the three most important computer vision tasks: object segmentation, object identification, and defect detection in video and still pictures.

Object Segmentation 

Object instance segmentation is the process of separating individual items in a warehouse’s containerized storage system. Instance segmentation guides subsequent robotic operations in object handling, including grip creation, motion planning, and placement. The success of picking, object recognition, and the introduction of flaws depend on the instance segmentation’s accuracy.

Object segments from over 50,000 photos have been manually labeled at a high-grade level. Instance segmentation algorithms face a new problem when dealing with variations in object types and clutter levels.

Object Identification 

The process of correctly assigning an image region to one of the items in a database is known as object identification (ID). This work can be done before or after a robot picks up an object. During the pre-pick phase, locating a portion of an item in the bag provides ready access to any previously acquired models or properties of the item for manipulation planning.

For robotic manipulation, this poses a problem of open set object recognition and confidence estimate. The dataset, which includes more than 190,000 distinct objects in various configurations, will evaluate and compare several picture retrieval and few-shot classification approaches that incorporate uncertainty estimates.

Defect Detection 

The defect identification task is determining whether or not a flaw was introduced during robotic manipulation. The dataset contains two types of robot-induced defects: multi pick and package-defect. A “package defect” occurs when the object’s packaging is damaged somehow, either because it was opened or broken into several pieces. When many items need to be moved from one container to another, this process, known as a “multi-pick,” is employed.

Rare but expensive robot-induced faults, such as those seen during multi-object pick and packaging, are detected using human-issued labels. The collection includes over 100,000 actions with no flaws and over 19,000 photos, and 4,000 videos of activities with problems.

In conclusion

Automating processes in contemporary warehouses requires a robotic manipulator to cope with a broad range of products, unstructured storage, and dynamically changing inventory. In such environments, perceiving the identity, physical properties, and status of objects during manipulation is difficult. There must be more diversity in item attributes, clutter, and interactions in existing robotic manipulation datasets since they only consider a small subset of objects or rely on 3D models to build synthetic settings. Using a robotic manipulator to conduct object singulation from containers containing diverse contents, researchers present a large-scale dataset obtained in an Amazon warehouse. Images, videos, and information representing over 235K pick-and-place operations on 190K distinct items are available in ARMBench. Information is recorded before the pick, during the transfer, and after placement. Learning generalizable representations that may transfer to other visual perception tasks will be made possible by the availability of large amounts of sensor data and fine-grained properties of objects. In addition, researchers want to add 3D data and annotations to the dataset and suggest more benchmarking activities.


Check out the Paper and Project. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our 18k+ ML SubRedditDiscord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.

🚀 Check Out 100’s AI Tools in AI Tools Club

🐝 Join the Fastest Growing AI Research Newsletter Read by Researchers from Google + NVIDIA + Meta + Stanford + MIT + Microsoft and many others...