Covariant

What Covariant is

Covariant is a company that works on artificial intelligence for robots. It was founded in 2017, and its founders and core managers mostly came from UC Berkeley^[1]. The company is known for building robotics foundation models: large models trained on broad data and then adapted to many different robot tasks, instead of being written for one narrow job.

The general idea behind a foundation model is to train once on a large body of data and reuse the result across many downstream tasks. Applied to robots, this is harder than applied to text, because the data comes from machines interacting with the physical world rather than from documents on the web. A robotics foundation model has to cope with objects, surfaces and physics that vary from one workplace to the next. That is the technical territory Covariant works in, and it is why its models are of interest to a company that runs large warehouses and logistics operations.

Founders and leadership

According to reporting by the Taiwanese financial news outlet Cnyes, Peter Chen was Covariant's co-founder and chief executive, and Rocky Duan was its chief technology officer^[1]. Both had previously been research scientists at OpenAI from 2016 to 2017^[1]. Tianhao Zhang became a co-founder of the company in 2017^[1]. Pieter Abbeel is also named among Covariant's founders^[1].

The 2024 transaction with Amazon

Amazon announced on 30 August 2024 that it would hire Covariant founders Pieter Abbeel, Peter Chen and Rocky Duan, together with about a quarter of the company's employees^[1]. Alongside the hiring, Amazon signed a non-exclusive licence to use Covariant's robotics foundation models^[1]. Because the licence is non-exclusive, it does not give Amazon sole rights to those models.

Covariant itself was not simply absorbed. It would continue operating under chief executive Ted Stinson, with Tianhao Zhang remaining at the company^[1]. So the founders and part of the team moved to Amazon, while the company stayed in existence as a separate business with a new chief executive and at least one of its co-founders still inside.

A "reverse acqui-hire"

Cnyes described the transaction as a reverse acqui-hire, similar to Amazon's earlier hiring of Adept's founders^[1]. An acqui-hire in the ordinary sense means buying a company mainly to obtain its people rather than its products: the technology is often secondary, and the point is to bring a proven team inside the buyer. Calling this one "reverse" points at the shape of the deal — the people were hired and the technology was licensed, but the company was left standing and continued to run its own business.

Such arrangements have become a common way for large technology companies to bring in AI talent without acquiring a whole company. For the startup, the deal can supply resources and a large customer; for the buyer, it brings a team quickly and gives access to models through a licence rather than through ownership.

Where it sits

Covariant belongs to a group of companies that formed around the idea that machine learning for robots should follow the same path as machine learning for language and vision: build a general model from large amounts of data, then adapt it. The 2024 deal placed its founding team inside Amazon, one of the largest operators of warehouses and logistics networks, while leaving the company itself independent and free to keep developing and licensing its models^[1]. For readers following how AI talent and AI models move between startups and large platforms, the transaction is a compact example: the people went one way, a licence went the other, and the original company kept operating.