What Is Covariant?
Covariant is a US-based robotics-AI company that builds software brains for warehouse picking arms rather than the arms themselves. Its core product, the Covariant Brain, is deployed on robotic arms from third-party manufacturers to handle item-picking tasks like each-picking, induction, and depalletizing in distribution centers.
The company was founded in 2017 by Peter Chen, Pieter Abbeel, and Rocky Duan, all with backgrounds in AI research including OpenAI and UC Berkeley. It built a reputation for training on real, high-volume warehouse pick data rather than only simulation, which became the foundation for its later foundation-model work.
This page covers what Covariant's technology does, its flagship RFM-1 model, the 2024 Amazon deal that reshaped the company, and where Covariant stands today.
| Attribute | Detail | Note |
|---|---|---|
| Founded | 2017 | Berkeley, California |
| Founders | Peter Chen, Pieter Abbeel, Rocky Duan | Joined Amazon in Aug 2024 |
| Flagship model | RFM-1 (Robotics Foundation Model) | Introduced March 2024 |
| Product | Covariant Brain | AI software for warehouse picking arms |
| Status | Operating independently | Non-exclusive AI license held by Amazon |
RFM-1 And The Covariant Brain
RFM-1, short for Robotics Foundation Model, is Covariant's flagship AI model, introduced in March 2024. Covariant describes it as an 8-billion-parameter multimodal transformer trained on text, images, video, robot sensor readings, and motor actions, treating all of them as tokens in one sequence so a robot can reason about a task the way a language model reasons about text.
The stated goal is to move picking robots away from narrow, single-task programming and toward a general-purpose model that can adapt to new items and environments with far less task-specific engineering. Covariant has since published updates describing in-context learning, where a deployed robot can adjust its grasping behavior on the fly after a failed attempt, without a full retraining cycle.
The Covariant Brain is the deployed product built on this research: a perception-and-control layer that plugs into robotic arms already installed in a warehouse, rather than a robot Covariant manufactures and sells itself. That software-first approach is why Covariant is usually classified alongside other robotics foundation-model builders rather than warehouse hardware brands.
- RFM-1: an 8B-parameter multimodal transformer for robot perception and action.
- Trained on text, images, video, sensor data, and motor actions as one token stream.
- In-context learning update lets deployed robots adapt after a failed grasp.
- Covariant Brain is the software layer, deployed on third-party robotic arms.
The 2024 Amazon Deal
On August 30, 2024, Amazon announced it had hired Covariant's three founders, Peter Chen, Pieter Abbeel, and Rocky Duan, along with roughly a quarter of Covariant's staff, in what reporting described as a reverse-acquihire structure. Amazon did not acquire Covariant outright; instead, it entered a non-exclusive licensing agreement for Covariant's robotic foundation models.
Covariant said the arrangement let it "optimize" its position while it continued serving its own customers, which at the time reportedly included the Otto Group, logistics firm Radial, and pharmaceutical distributor McKesson. Ted Stinson, previously Covariant's COO, moved into the CEO role to lead the company after the founders departed.
This kind of deal, where a large company hires a startup's leadership and licenses its technology instead of buying the company, has become a recognizable pattern in AI. It let Amazon absorb the research talent and a usable model license quickly, while leaving Covariant as a going concern rather than winding it down.
Covariant's Status Today
As of 2026, Covariant continues to operate as an independent company under CEO Ted Stinson, rather than being folded into Amazon. The August 2024 agreement was explicitly non-exclusive, meaning Covariant retains the right to license and develop its models for other warehouse customers even though Amazon also holds a license.
Because Covariant sells software rather than a branded robot, its market position is easiest to compare against other robotics foundation-model builders and warehouse automation vendors rather than against a single physical product. It sits in the same broader category as the picking-robot vendors covered in the warehouse picking robots guide and the AI-model builders covered alongside it in the Physical Intelligence company profile.
The honest takeaway for anyone researching Covariant is that the brand survived the loss of its founders, but its post-2024 roadmap, customer roster, and next model release are set by a smaller team than the one that built RFM-1. Treat any claim about Covariant's current deployments as something to verify against the company's own recent announcements rather than 2024-era coverage.
How Covariant Fits Among Warehouse AI Vendors
Covariant is one of several companies building AI specifically for item picking and manipulation in warehouses, a narrower slice of the broader warehouse automation market. Its closest comparison point is any vendor supplying the perception-and-control software for picking arms rather than the arms or the facility-wide fleet software.
For the wider hardware landscape those AI models plug into, see the warehouse picking robots guide and the autonomous mobile robots guide for the fleet side of a distribution center. For coverage of other robotics foundation-model builders working on the same general-purpose-robot-brain problem, see the Physical Intelligence profile.
Buyers evaluating any of these AI-first vendors should ask the same question the founders' departure raises directly: who on the current team owns the model roadmap, and what is the actual deployment count today rather than at last announcement.
Bottom Line
Covariant built RFM-1, a foundation model meant to give warehouse picking robots language-model-style reasoning over vision and action, and packaged it as the Covariant Brain for arms already installed on warehouse floors. In August 2024, Amazon hired its three founders and a quarter of its staff and took a non-exclusive license to its models, and Covariant kept operating under new CEO Ted Stinson rather than shutting down. Anyone researching Covariant today should treat its founder-era coverage as history and verify current customers and roadmap directly.
Compare Covariant's approach against the physical picking hardware it plugs into in the warehouse picking robots guide, and see how it stacks up against other robotics foundation-model builders in the Physical Intelligence company profile.
FAQs
What does Covariant do?
Covariant builds AI software, most notably its RFM-1 foundation model and the Covariant Brain product, that gives warehouse picking robots the ability to perceive items and plan grasps across a wide range of products rather than being programmed for one task at a time.
What is RFM-1?
RFM-1, Covariant's Robotics Foundation Model, is an 8-billion-parameter multimodal transformer introduced in March 2024, trained on text, images, video, sensor data, and motor actions to let robots reason about physical tasks.
Did Amazon acquire Covariant?
No. In August 2024, Amazon hired Covariant's three founders and about a quarter of its staff and took a non-exclusive license to its AI models. Covariant was not acquired and continued operating as an independent company.
Is Covariant still in business?
Yes. Covariant continues to operate independently under CEO Ted Stinson, who moved from COO into the role after the founders left for Amazon in 2024.
Who founded Covariant?
Covariant was founded in 2017 by Peter Chen, Pieter Abbeel, and Rocky Duan. All three joined Amazon in August 2024 as part of a licensing and hiring agreement, though the company they founded kept operating.