CtrlK
| Covariant |
|---|
| Covariant, Inc. |
| Type |
| Industry |
| Founded |
| Founders |
| Headquarters |
| Key people |
| Products |
| Total funding |
| Valuation |
| Website |
Covariant, Inc. (originally Embodied Intelligence) is an American artificial intelligence company that develops AI software for robotic manipulation and warehouse automation. Founded in October 2017 by UC Berkeley professor Pieter Abbeel and three of his graduate students, the company builds the Covariant Brain, an AI platform that enables robots to perceive, reason, and act autonomously in unstructured environments such as distribution centers and fulfillment warehouses.[1][2] In March 2024, Covariant introduced RFM-1, described as the first commercial robotics foundation model, an 8-billion-parameter multimodal transformer trained on text, images, video, robot actions, and sensor data.[3]
The company raised $222 million in total venture funding and was valued at $625 million as of its final funding round in April 2023.[4] In August 2024, Amazon announced a licensing agreement for Covariant's technology and hired three of the four co-founders along with roughly one quarter of the company's workforce to join Amazon's Fulfillment Technologies and Robotics division.[5]
In October 2017, Pieter Abbeel and three of his graduate students at the University of California, Berkeley, Peter Chen, Rocky Duan, and Tianhao Zhang, co-founded a company called Embodied Intelligence.[1][6] All four founders had backgrounds in reinforcement learning and robot learning research at Berkeley's Robot Learning Lab and the Berkeley Artificial Intelligence Research (BAIR) Lab. Abbeel, Chen, and Duan had also worked as researchers at OpenAI before starting the company.[2][7]
The original concept behind Embodied Intelligence was to develop AI software that would allow robots to learn new tasks through virtual reality (VR) demonstrations. A human operator wearing a VR headset would perform a task, and the robot would learn to replicate it using a combination of imitation learning and reinforcement learning.[6] The company raised $7 million in seed funding in November 2017, led by Amplify Partners with participation from Lux Capital, SV Angel, FreeS, 11.2 Capital, and A.Capital.[6][8]
The company operated in stealth mode for roughly two and a half years, during which the team refined its approach and developed the core technology that would become the Covariant Brain. The founders shifted their focus from VR-based teaching to building a universal AI platform for robotic manipulation in real-world industrial settings. By 2018, the Covariant Brain had achieved what the company described as "human-level autonomy" in certain picking tasks.[9]
During this stealth period, the company also raised a $20 million Series A round in the first half of 2019. Amplify Partners led the round, and notable angel investors included Geoffrey Hinton, Yann LeCun, Jeff Dean, and Raquel Urtasun.[2][10]
Covariant officially launched from stealth on January 29, 2020, revealing its new name and the Covariant Brain platform.[2] At the time of launch, the company had already deployed robotic stations at customer facilities in North America and Europe. One of the earliest production deployments was at Obeta, a German electrical supply wholesaler outside Berlin, where the Covariant Brain powered a KNAPP Pick-it-Easy Robot for automated goods-to-person order picking.[2][10]
Peter Chen served as CEO, Pieter Abbeel as President and Chief Scientist, and Rocky Duan as Chief Technology Officer.
In February 2020, shortly after emerging from stealth, Covariant announced a partnership with ABB, the Swiss-Swedish multinational and one of the world's largest suppliers of industrial robots. ABB had conducted a global competition in 2019, evaluating 20 AI technology startups on 26 real-world picking, packing, and sorting challenges. Covariant won the competition, and the two companies agreed to co-develop AI-enabled robotic solutions for warehouse automation.[11] The first deployment from this partnership was installed at Active Ants, an e-commerce fulfillment provider (part of the bpost group) in Utrecht, Netherlands.[11]
In March 2020, Covariant also announced a partnership with KNAPP, an Austrian logistics automation company. KNAPP integrated the Covariant Brain into its Pick-it-Easy Robot system, combining KNAPP's KiSoft software and mechanical hardware with Covariant's AI-driven perception and manipulation capabilities.[12] The KNAPP partnership proved particularly productive: over the following five years, it resulted in deployments across more than 26 unique customers in Europe, the Americas, Asia, and Australia.[13]
On May 6, 2020, Covariant announced a $40 million Series B funding round led by Index Ventures, with participation from Radical Ventures and existing investor Amplify Partners. Mike Volpi from Index Ventures joined the company's board of directors. This brought total funding to $67 million.[14] The capital was used to expand into new industries and geographies. By this point, the company reported that its AI-powered robots were running consistently at customer facilities, with mean unassisted operating time exceeding one hour.[14]
In July 2021, Covariant raised $80 million in a Series C round led by Index Ventures, with participation from Amplify Partners, Radical Ventures, and new investors Temasek and Canada Pension Plan Investment Board (CPP Investments). This brought total capitalization to $147 million.[15] The company used the funds to expand its engineering and research teams and grow its global deployment footprint.
The company continued to expand its product line, introducing solutions for depalletization and kitting in 2022, and opening a London office that same year.[9]
In April 2023, Covariant raised an additional $75 million in a Series C extension, bringing total funding to $222 million and valuing the company at $625 million. Radical Ventures and Index Ventures co-led the round, with participation from returning investors CPP Investments and Amplify Partners, and new investors Gates Frontier Holdings, AIX Ventures, and Northgate Capital.[4] CEO Peter Chen noted that the company had experienced 6x revenue growth in 2022.[4]
On March 11, 2024, Covariant introduced RFM-1 (Robotics Foundation Model 1), which it called the first commercial foundation model for robotics. RFM-1 was designed to give robots reasoning capabilities comparable to the way large language models process text, but applied to physical-world interactions.[3] (See the Technology section below for technical details.)
On August 30, 2024, Amazon announced that it had entered into a licensing and hiring agreement with Covariant. Under the terms of the deal, Amazon received a non-exclusive license to Covariant's robotic foundation models. Three of the four co-founders, Pieter Abbeel, Peter Chen, and Rocky Duan, along with approximately 25% of Covariant's workforce, joined Amazon's Fulfillment Technologies and Robotics team.[5][16]
Covariant's COO, Ted Stinson, assumed the role of CEO, and co-founder Tianhao Zhang remained with the company to continue leading technical development. Covariant stated it would continue serving its existing customers and developing its AI robotics platform independently.[17]
The deal drew comparisons to similar arrangements between other large technology companies and AI startups. The Verge and other outlets described the structure as a "reverse acquihire," in which a technology giant hires key personnel and licenses technology without formally acquiring the company, potentially avoiding the antitrust scrutiny that would accompany a traditional acquisition.[5][16]
In January 2025, a whistleblower filed a complaint with the Federal Trade Commission (FTC), Securities and Exchange Commission (SEC), and Department of Justice (DOJ), alleging that the deal was structured to circumvent premerger reporting requirements. The complaint claimed the total value of the arrangement was approximately $380 million, with an additional $20 million licensing payment due one year after closing, a figure that exceeded the $119.5 million threshold for mandatory premerger filings under the Hart-Scott-Rodino Act. The whistleblower also alleged that certain terms of the deal restricted Covariant's ability to license its technology to other companies.[18] The FTC confirmed it was reviewing the complaint but declined to comment further.[18]
In December 2025, Amazon appointed Pieter Abbeel to lead its frontier model research team within the AGI organization, while he continued his work on Amazon's robotics efforts.[19]
The Covariant Brain is the company's core AI platform, designed to serve as a universal intelligence layer for robotic manipulation. The system combines deep learning, computer vision, reinforcement learning, and imitation learning to allow robots to perceive their environment, identify and understand objects, plan movements, and execute grasping and placing tasks in real time.[2][14]
Unlike traditional industrial robot programming, which requires objects to be pre-defined and environments to be precisely controlled, the Covariant Brain enables robots to handle previously unseen objects in unpredictable configurations. The system uses 3D perception to understand object shapes, physical affordance reasoning to determine how best to grasp each item, and few-shot learning to adapt quickly to new product categories.[14]
A distinctive feature of the Covariant Brain is its fleet learning architecture. Every robot running the Covariant Brain shares data with a central model, so each individual robot benefits from the collective experience of the entire deployed fleet. Covariant has described this as robots "learning from millions of picks performed by connected robots in warehouses around the world," allowing continuous improvement across all deployments.[20]
As of 2023, the Covariant Brain had been deployed across industries including fashion and apparel, health and beauty, industrial supply, pharmaceuticals, grocery, parcel handling, and general merchandise.[4][15]
RFM-1 is an 8-billion-parameter transformer model trained on a combination of internet-scale data (text and images) and what Covariant described as "the largest real-world robot production dataset" collected from its deployed fleet. The training data encompasses multiple modalities: visual data from multiple camera angles, video sequences, task descriptions in natural language, numerical sensor readings from motor encoders and pressure sensors, and performance metrics from production operations running at 1,000+ cycles per hour with 99%+ precision.[3]
The model functions as a multimodal any-to-any sequence model. All input modalities are tokenized into a shared representation space, and the model performs autoregressive next-token prediction across these modalities. This means RFM-1 can accept any combination of text, images, video, actions, and sensor data as input, and generate any combination of these modalities as output.[3]
Key capabilities of RFM-1 include:
| Capability | Description |
|---|---|
| Physics simulation | Functions as a world model by predicting future video frames given initial images and planned robot actions; learns both frame-by-frame dynamics and high-level outcomes (for example, predicting the state of a bin after a grasping attempt) |
| Language understanding | Processes natural language instructions, enabling operators to direct robots in plain English without traditional programming; robots can also request assistance and suggest strategies |
| In-context learning | Adapts to new tasks or environments based on a small number of examples provided at inference time, without requiring retraining |
| Video generation | Generates AI-produced videos showing predicted outcomes of robotic actions, allowing the system to simulate future scenarios and select the best course of action |
At the time of its announcement, Covariant acknowledged several limitations of RFM-1: it had not yet been deployed to production customers, it operated at relatively low resolution (approximately 512x512 pixels) and frame rate (approximately 5 fps), its limited context length made it difficult to model very small objects or very rapid motions, and its orchestration logic still relied on traditional programming rather than the foundation model itself.[3]
Covariant's technology has been deployed across several distinct warehouse automation use cases:
| Application | Description | Performance |
|---|---|---|
| Goods-to-Person Picking | Robots select individual items from bins or totes and place them into order containers as part of an automated goods-to-person fulfillment workflow | Deployed across fashion, health and beauty, pharmaceuticals, grocery, and industrial supply sectors |
| Robotic Putwall | Robots sort batch-picked items into individual customer orders at putwall stations | Up to 515 picks per hour (PPH) with less than 0.1% of orders requiring human intervention[21] |
| Robotic Induction (Apparel) | Automates induction of polybagged apparel onto unit sorters (split tray, tilt tray, bomb bay), pocket sorters, AMRs, and autobaggers | 1,300+ units per hour (UPH)[20] |
| Robotic Induction (Parcel) | Dual-robot configuration for singulating and inducting flats and mixed parcels onto conveyor systems | 2,800+ UPH[20] |
| Depalletization | Robots remove items or cases from pallets for downstream processing | Introduced in 2022[9] |
| Kitting | Robots assemble groups of related items into kits for packaging or distribution | Introduced in 2022[9] |
Covariant integrates with hardware from multiple robotics and automation partners, including ABB robotic arms and KNAPP's Pick-it-Easy Robot system. The company also supports integration with sorter systems from Beumer, Eurosort, and AMR-based sorters such as the Tompkins tSort.[20]
Covariant has deployed its technology at warehouses and distribution centers across Europe, North America, and the Asia-Pacific region. Notable customers include:
| Customer | Industry | Deployment Details |
|---|---|---|
| KNAPP | Logistics automation | Primary integration partner; Pick-it-Easy Robot powered by Covariant Brain deployed at 26+ customer sites across four continents[13] |
| Otto Group | E-commerce/Retail | Long-term strategic partnership announced May 2023; initial deployments at Haldensleben and Altenkunstadt facilities in Germany, with a vision of deploying hundreds of robots across all Otto Group fulfillment centers[22] |
| GXO Logistics | Third-party logistics (3PL) | Deployed KNAPP Pick-it-Easy Robot for automated pocket induction at a fashion e-commerce warehouse in Tilburg, Netherlands[23] |
| McKesson | Pharmaceutical distribution | Deployed KNAPP Pick-it-Easy Robot for picking small pharmaceutical items such as pill bottles; runs around the clock to address labor shortages[24] |
| Capacity LLC | Third-party logistics (3PL) | Five Covariant Robotic Putwalls sorting batch-picked orders for health and beauty brands; first station went live in June 2021[21] |
| Radial, Inc. | E-commerce fulfillment | Deployed 12 AI-powered robotic putwalls[4] |
| Wurth Industrie Service | Industrial supply | Deployed KNAPP Pick-it-Easy Robot for automated order picking from goods-to-person shuttle system[25] |
| Brodrene Dahl | Technical wholesale | Norwegian supplier modernizing warehouse operations with AI-powered picking[13] |
| BESTSELLER | Fashion retail | Global fashion retailer using the system in a new distribution facility[13] |
| Active Ants (bpost) | E-commerce fulfillment | First ABB-Covariant deployment in Utrecht, Netherlands[11] |
| Obeta | Electrical wholesale | Early deployment site in Germany; one of the first Covariant Brain production installations[2] |
Covariant raised a total of $222 million across five funding rounds:
| Round | Date | Amount | Lead Investor(s) | Key Participants |
|---|---|---|---|---|
| Seed | November 2017 | $7 million | Amplify Partners | Lux Capital, SV Angel, FreeS, 11.2 Capital, A.Capital[6][8] |
| Series A | H1 2019 | $20 million | Amplify Partners | Geoffrey Hinton, Yann LeCun, Jeff Dean, Raquel Urtasun (angels)[2][10] |
| Series B | May 2020 | $40 million | Index Ventures | Radical Ventures, Amplify Partners[14] |
| Series C | July 2021 | $80 million | Index Ventures | Amplify Partners, Radical Ventures, Temasek, CPP Investments[15] |
| Series C Extension | April 2023 | $75 million | Radical Ventures, Index Ventures | CPP Investments, Amplify Partners, Gates Frontier Holdings, AIX Ventures, Northgate Capital[4] |
The April 2023 round valued the company at $625 million.[18]
Pieter Abbeel (President and Chief Scientist, 2017-2024) is a Belgian-American computer scientist who has served as a professor at UC Berkeley's Department of Electrical Engineering and Computer Sciences since 2008. He directs the Berkeley Robot Learning Lab and co-directs the Berkeley Artificial Intelligence Research (BAIR) Lab. His research focuses on deep reinforcement learning, imitation learning, meta-learning, and transfer learning. Before co-founding Covariant, Abbeel worked as a researcher at OpenAI and co-founded Gradescope (acquired by Turnitin in 2018). He received his Ph.D. from Stanford University in 2008. His honors include the ACM Prize in Computing, the NSF CAREER Award, and the IEEE Fellow designation. He joined Amazon in August 2024 and was named head of Amazon's frontier model research within the AGI organization in December 2025.[7][19]
Peter Chen (CEO, 2017-2024) was a graduate student under Abbeel at UC Berkeley and a former researcher at OpenAI. He joined Amazon in August 2024 as part of the licensing and hiring agreement.[5]
Rocky Duan (CTO, 2017-2024) was a graduate student at UC Berkeley and a former researcher at OpenAI. He joined Amazon alongside Chen and Abbeel in August 2024.[5]
Tianhao Zhang (co-founder) was a graduate student at UC Berkeley and a former researcher at Microsoft. He remained with Covariant after the Amazon deal, taking on a leadership role alongside CEO Ted Stinson.[17]
Ted Stinson became CEO in September 2024 after previously serving as Covariant's Chief Operating Officer. He and Tianhao Zhang lead the company's ongoing operations, customer relationships, and technology development.[17]
Covariant operates in the AI-powered warehouse robotics market, which includes several other companies developing machine learning-driven solutions for picking, packing, sorting, and other logistics tasks:
Covariant differentiated itself from these competitors primarily through its fleet learning approach (where all deployed robots contribute data to a shared model) and its early investment in building a foundation model for robotics (RFM-1).[3][4]