# Covariant

> Covariant built an AI layer for warehouse robots that could learn to pick unfamiliar products rather than be programmed SKU by SKU. Founded in 2017 by UC Berkeley and OpenAI researchers, it turned real warehouse deployments into a data flywheel, expanded from picking into sortation and induction, and unveiled the multimodal RFM-1 foundation model in 2024. Months later Amazon hired three founders and roughly a quarter of the team while taking a non-exclusive license to Covariant's models; the remaining company said it would continue serving dozens of customers.

- **Founded:** 2017
- **Headquarters:** Emeryville, United States
- **Founders:** Peter Chen (Co-founder and former CEO), Pieter Abbeel (Co-founder and former President and Chief Scientist), Rocky Duan (Co-founder and former CTO), Tianhao Zhang (Co-founder and Head of Product)
- **Products:** Covariant Brain, Robotic Putwall, Robotic Induction, RFM-1
- **Notable:** Moved from stealth to a production deployment at Obeta in Germany, where its system handled glossy, reflective and plastic-bagged products., Built a fleet-learning system in which deployments contributed experience to a common model., Radial reported about 425 puts per hour and roughly 100,000 picks per robot per month at full operation for its robotic putwalls.

## Products & services

- **Covariant Brain** — A shared AI control platform combining perception, motion planning, imitation learning and reinforcement learning so industrial arms can manipulate unfamiliar objects.
- **Robotic Putwall** — An integrated sortation station that picks items from mixed-SKU totes, scans them and places them into order cubbies while connecting to warehouse-management and pick-to-light systems.
- **Robotic Induction** — An AI-powered station that places varied items onto conveyors or sorters for high-throughput fulfillment workflows.
- **RFM-1** — A multimodal robotics foundation model that accepts text, images, video, robot actions and sensor data, and can generate actions, language or predicted video outcomes.

## Achievements

- Moved from stealth to a production deployment at Obeta in Germany, where its system handled glossy, reflective and plastic-bagged products.
- Built a fleet-learning system in which deployments contributed experience to a common model.
- Radial reported about 425 puts per hour and roughly 100,000 picks per robot per month at full operation for its robotic putwalls.
- KNAPP reported Covariant-powered Pick-it-Easy robots in use at 26 customers across three continents.
- Secured an Otto Group program beginning with more than 100 robots and designed to expand across its fulfillment network.
- Released RFM-1 in March 2024, extending the company's warehouse data into a multimodal robotics foundation model.
- Licensed its robotic foundation models to Amazon in August 2024 while three founders and about one-quarter of staff joined Amazon Robotics.

## Latest updates

- **2023-04** — Raised a $75 million Series C extension to scale deployments and research.
- **2023-05** — Otto Group announced a strategic partnership beginning with more than 100 Covariant-powered robots in Germany.
- **2024-03** — Introduced RFM-1, a multimodal robotics foundation model trained on robot and internet data.
- **2024-08** — Amazon licensed Covariant's models non-exclusively and hired Pieter Abbeel, Peter Chen, Rocky Duan and roughly a quarter of the team; Covariant said it would continue serving customers.
- **2026-09** — Covariant's public website presents a minimal holding page; its major public corporate update remains the August 2024 Amazon agreement.

## Links

- Website: https://covariant.ai/
- LinkedIn: https://www.linkedin.com/company/covariant-ai
- Twitter/X: https://twitter.com/CovariantAI
- YouTube: https://www.youtube.com/@covariantai

---

Profile page: https://yespress.io/covariant
Published by YesPress — https://yespress.io
Last updated: 2026-09-28