In 2012, Amazon discreetly purchased a robotics company named Kiva Systems, a step that significantly boosted the efficiency of its ecommerce operations and initiated a broader shift towards warehouse automation.
Last week, the ecommerce behemoth announced another transaction that could be equally transformative, agreeing to acquire the founders of Covariant, a startup exploring how AI can automate the sorting and handling of various physical items.
Covariant’s journey in commercializing AI-powered industrial robots might have been tough due to high expenses and intense market competition; however, the agreement also allows Amazon to utilize Covariant’s models and data, potentially leading to another profound shift in ecommerce, one that might be difficult for competitors to rival given Amazon’s extensive operational scale and data assets.
This deal is also indicative of a larger trend where major tech companies acquire essential talent and know-how from AI startups without owning the companies outright. Amazon followed a similar path with startup Adept in June. Microsoft in March entered into a partnership with Inflection, and in August, Google recruited the founders of Character AI.
During the early 2000s, Kiva introduced an innovative method to transport products within warehouses by using short robots to transport shelves filled with goods directly to human workers, preventing the need for extensive walking in search of items. These mobile robots, which bear resemblance to those used in manufacturing, operate under sophisticated algorithms to manage numerous robots coexisting within the same workspace.
By 2023, Amazon had expanded its fleet of mobile robots from about 10,000 in 2013 to 750,000. This massive deployment allowed the company to deliver millions of items with greater speed and cost-efficiency than its competitors.
As reported by WIRED, in recent years, Amazon has adopted advanced robotic technologies that utilize machine learning for tasks like identifying, grabbing, and sorting packages. These developments have been propelled forward by the vast amount of data generated as goods move through Amazon’s facilities, enhancing the capabilities of their algorithms. This advancement has significantly automated processes previously handled by humans in several of its fulfillment centers.
However, a persistently challenging task for automation is the physical handling of products. This process demands flexibility to manage variables like friction and potential slipping, with the robots frequently encountering new or irregularly shaped items in Amazon’s extensive inventory.
Covariant has been focusing on developing AI algorithms capable of handling a diverse range of items reliably over the past few years. The company, established in 2020, was founded by Pieter Abbeel, a professor at UC Berkeley renowned for his innovative work in the application of machine learning to robotics. Alongside him were his students, including Peter Chen, who took on the role of CEO, and Rocky Duan, named CTO. Recently, a deal was arranged for all three founders, as well as several research scientists from the start-up, to join Amazon.
According to Alexandra Miller, a spokesperson from Amazon, “Covariant’s models will now power some of the robotic manipulation systems used throughout our fulfillment network.” Amazon has chosen not to disclose the financial terms of this acquisition.
Pieter Abbeel, who was also an early employee at OpenAI, drew inspiration for his company from the success story of ChatGPT. In March, Covariant showcased a chat interface for its robot and mentioned that they had developed a foundation model for robotic grasping, which is an algorithm designed to improve with increased data input. At that time, CEO Chen highlighted the primary challenge as acquiring sufficient data to enhance the capabilities further. Considering the massive inventory in Amazon’s warehouses and the scope of their operations, Amazon appears to be an ideal match to provide such data.
Yaro Tenzer, cofounder and CEO of RightHand Robotics, a company based in Boston that provides robotic picking systems, noted that startups like Covariant often face difficulties in delivering cost-effective solutions due to the complexities of AI development, customization demands for different projects, and lengthy sales cycles associated with ecommerce fulfillment.
“Building a sustainable hardware and software business that justifies venture capital growth is extremely challenging,” Tenzer says. That said, Tenzer speculates that Amazon may have seen an opportunity to build a new team dedicated to AI-powered picking, although he believes that would most likely take several years to pay off.
Whether the Covariant deal turns out to be as pivotal as the Kiva one remains to be seen. But when WIRED visited one of Amazon’s more automated facilities last year, humans were still required for picking and packing—as well as for helping the odd lost or confused robot. The Amazon–Covariant deal could well see the numbers of humans needed to handle its products dwindle in the decade to come.
