Enhancing Pick-and-Place Robots with Annotations from Scale Rapid

Overview

High-Quality Annotations are a Vital Part of Rapidly Improving Models

Ambi Robotics provides customers with AI-powered robotic systems that empower them to scale their operations and handle increasing supply chain demand. Logistical warehouses across the e-commerce industry depend on this technology to automate the process of picking up packages, scanning their barcodes, and sending them to the correct location. The machine learning (ML) system is responsible for not just identifying an object and its location, but also for moving the robot hand to that location to grasp the object. For these customers, pick success rate – how often a robot successfully picks up an object – is the most important marker of success. Honing these models so that they provide the best possible performance to each client typically requires the collection and annotation of new data. To accomplish this, Ambi Robotics needed a scalable solution for obtaining high-quality annotations. Today, Ambi Robotics is working with Scale Rapid to get the data annotations that they need to continuously enhance their pick-and-place models.

The Problem

Ambi Robotics Needed a Scalable Approach for Data Annotation

Training a well-performing ML model to perform a task as complex as pick-and-place requires a lot of annotated data. While initially Ambi Robotics was managing the annotation process in-house, the approach was simply not scalable for the amount of data they needed. When working with new clients and locations, for example, Ambi Robotics would sometimes see lower pick-and-place success rates, simply because the environment looked different. The best way to improve performance was to mine data from the new location, annotate it, and then retrain their ML model.

The company didn’t have the infrastructure to process this large quantity of data on a recurring basis, however – Regularly generating thousands of labels in 24-48 hours internally wasn’t a feasible solution. To provide all of their customers with the best possible experience, they needed a more scalable approach to data annotation.

Before we started working with Scale Rapid, we were annotating data in-house. We didn’t have the infrastructure to complete thousands of annotations quickly, so we needed to find a scalable approach for data annotation. With Scale Rapid, we can request annotations and get results within 24 to 48 hours.
Yi Li
Head of Computer Vision
Ambi Robotics

The Solution

With Scale Rapid’s Tooling and Infrastructure, Ambi Robotics Obtains High-Quality Annotations Quickly

Now, when Ambi Robotics works with a new client, they can outsource the data annotation process to Scale Rapid, quickly obtaining the annotations they need to improve their ML models. Scale Rapid offers a simple step-by-step approach for setting up the annotation process, making it easy for customers to set up a new task in just minutes.

In order to provide high-quality data, Scale Rapid builds quality evaluation into their pipeline through batch annotation. Scale Rapid produces a calibration dataset to identify gaps in the labeling instructions, until the customer is satisfied with the data calibration. After calibration is complete, Scale Rapid produces a production batch for the client.

Scale Rapid has also provided Ambi Robotics with a reduction in lead time – While Ambi Robotics’ internal annotation process could sometimes take weeks, Scale Rapid consistently delivers production-ready results within 1-2 days. When assessing data annotation quality, Ambi Robotics has found that the results delivered by Scale Rapid are consistently on-level with their in-house annotation, while being produced using robust and scalable infrastructure.

"Scale Rapid’s annotations not only excel in terms of turnaround time, price, and accuracy, but the service also provides a unique annotation process that is easy to configure. Scale has very specific quality control tasks that are performed over batch annotation, and it automatically sets up training and validation tasks. This pipeline makes it easy to automatically tune data quality."
Yi Li
Head of Computer Vision
Ambi Robotics

The Result

Ambi Robotics Provides New Customers with Better Robotics Performance

The real world is filled with edge cases, so Ambi Robotics is always looking to increase their robotic systems’ pick success rate. Improving this metric provides customers with a better pick-and-place experience, and since Ambi Robotics began working with Scale Rapid, the model improvement process has become much easier.

Now, when the company sees degraded model performance at a new customer location, they can simply mine the relevant data and send it to Scale for annotation. Then, they can retrain their ML models using these data annotations, making the pick-and-place performance of their robotics systems more robust to new clients and locations.

After incorporating Scale Rapid into their production pipeline, Ambi Robotics can obtain high-quality data annotations quickly and easily, making it simple to use continuous learning to improve their system’s pick success rate.

"At Ambi Robotics, we perform a quality control task every time we roll out our robotic systems to new customers. Because new customers have different data distributions, we sometimes see pick success fall a bit. With Scale Rapid, we are able to mine data from the customer, send it to Scale for annotation, and then retrain the model with the new data added. This helps us to continuously improve our product."
Yi Li
Head of Computer Vision
Ambi Robotics