With the rise of online subscription boxes, a new kitting robot wants to help make things quicker.
With thousands of unique subscription services worldwide and more popping up every day, the online subscription service has been booming. Ambi Robotics has created an automated solution called AmbiKit to help e-commerce companies use artificial intelligence to learn how to identify and collect items out of bulk storage and separate them one by one.
The five-arm kitting solution boasts a picking line that is said to be 10,000 times faster than other alternatives. It can sort up to 60 items per minute in a commercial production environment and package kits such as online subscription boxes, shipping boxes and much more. Through the cloud, companies can teach the robot new grasping methods for different boxes and share those automatically down the production line.
Ken Goldberg of Ambi Robotics (formerly known as Ambidextrous Laboratories) has been working to solve the issue of universal picking since his undergraduate years at the University of Pennsylvania. While humans learn how to grasp different objects of various shapes and sizes at a young age, robots, on the other hand, cannot. It wasn’t until he was with one of his graduate students, Jeff Mahler—co-founder of Ambi Robotics—when they had a breakthrough to apply deep learning to the problem. The team researched into deep learning, physics, the mechanics of grasping, and combining it all with 3D sensing, depth sensors, and a neural network. They came across the idea to generate large datasets of robust grasps and apply algorithms to train the robot to work in real physical environments.
Since then, AmbiKit has raised over $6.1 million in seed funding for its picking robots and operating system that is based on simulation-to-reality AI. The machine is powered by Dex-Net, which rapidly teaches robots to pick and pack a variety of items by executing millions of cycles on a digital twin.
Recounting the beginning days of production, Goldberg jokes that they would ask people to throw a set of car keys at the robot. If the robot successfully picked them up, they could keep the car. (The robot would always pick up the keys.) In other cases, they would let people dump things in a clear box and allow the robot to sort them. The robot heavily relies on sensing, control and physics to sort items, meaning one small error could be the difference between dropping the object and picking it up successfully.
The Technology Behind Digital Retail
AmbiKit’s software runs on Swiss robotics company ABB’s off-the-shelf industrial machine. The system features a high-resolution 3D sensor and two arms, one with a conventional robot gripper and the other with a suction system. Both arms are controlled by a different neural network.
Using the software, the robot tries picking up objects in a virtual environment by using a deep neural network. The training process requires a lot of trial and error—a difficult task even in simulation. Dex-Net can catalogue items that have been seen before, to differentiate each object from unknown ones. It will also keep information on how to grasp an item, move an item to get a better look, or scan an object to decide whether to grab or suck up the object.
With a wide array of features, AmbiKit can complete order personalization without costing retailers millions of dollars annually due to human error, high employee turnover, and employee injury. Associates will be needed to work alongside the Ambikit system to load products and pack completed kits for shipping. Operators can select up to 15 SKUs for kitting by putting items in robot-directed areas.
According to Goldberg, it is not so much about eliminating warehouse workers, but rather cooperating with robots to reach the demand for labor. The process is simple. Workers place boxes of packed objects in bubble wrap in front of the robot. From there, the robot scans the package and puts it into the right bin—a method called sorting.
The robot can complete 12 kits per minute with 99 percent accuracy, says Ambi Robotics. It also requires zero training time for changeover or new items while offering real-time performance views.
As Goldberg puts it, Ambi Robotics is very customer-centric and needs to understand the customer’s problem in order to tailor a system that can deliver what is needed. This involves learning about new sensors, actuators, and software that can be reliable for the market. Throughout the entire process, the robot can help the customer, retailer, and parcel transportation company. AI can help e-commerce companies find a solution that can be deployable on the first day even in a location with less data bandwidth.
The Limitations of Universal Picking
On the other hand, universal picking has many limitations in robot perception and control. Sensor noises and occultation can obscure the geometry and positions of objects in the environment. It can also challenge the actuation and calibration of the arm positioning.
Universal picking uses a database of grasps on three-dimensional object models using grasp performance metrics with stochastic sampling. This requires a perception system, such as an industrial depth camera, to register sensor data for known objects. Machine learning can also be used to train function approximators such as deep neural networks to predict the probability of grasps, using images from training datasets.
To learn a policy, the software uses a training dataset based on physics and geometry to synthesize point clouds, grasps, and reward labels for heterogenous grippers. The synthetic training environment stimulates grasp outcomes by measuring resist forces and torques, and renders depth images of the 3D objects in bins using the camera position, focal length, and optical center pixel.
On the bright side, universal picking has the power to help speed up kitting in many scenarios such as warehouse, manufacturing, medicine, retail, and service—just like AmbiKit aims to do.
