Covariant

Snapshot

Covariant is a Berkeley-born startup that sold warehouse robot "brains" — a software layer that lets a conventional pick-and-place arm identify, grasp and place arbitrary items it has never seen. Between 2017 and 2024 it raised roughly $222M and earned a reputation as the most technically respected name in warehouse robotics AI, with deployments at Knapp, Obeta, Radial and Pitney Bowes. Then, on 29 August 2024, Amazon executed a "reverse acquihire": it hired the three co-founders and about a quarter of Covariant's staff and took a non-exclusive license to the company's foundation models, reportedly worth close to $400M in exit value to shareholders. What remains of Covariant is a hollowed-out shell that still services existing customers under CEO Ted Stinson. The page matters because it is the clearest case study of how AI robotics companies now actually "exit" — through regulatory-shy licensing deals that leave the IP in a hyperscaler's hands and the enterprise in limbo.

Founding story

Covariant (originally Embodied Intelligence) was spun out of UC Berkeley's robot-learning lab in late 2017 by Pieter Abbeel, Peter Chen, Rocky Duan and Tianhao Zhang. Abbeel — the senior partner — was a Berkeley professor who had co-authored the seminal 2010 paper on deep reinforcement learning for robot manipulation and had done a short stint as a research scientist at OpenAI. Chen, Duan and Zhang were his PhD students who had also rotated through OpenAI as early members of the research team. The group's thesis in 2017 was heretical for the robotics industry at the time: that progress in picking would come not from bespoke mechanical engineering but from scaling a single neural network across many different robots doing the same task, letting every grasp a Covariant-powered arm made improve the model for every other arm. The founders deliberately chose warehousing rather than consumer or home robotics, because repetitive enterprise picking gave them the volume of structured demonstration data a learning system needs to actually learn. The company stayed underground for two years, published a handful of research papers under the rebranded Covariant name in 2019, and formally emerged in early 2020 with its first paying customer, the German electrical wholesaler Obeta.

How it works

Every Covariant installation is a hardware-agnostic brain attached to someone else's arm. The company does not sell a robot. It sells an industrial-grade AI stack — the Covariant Brain — that integrates with an existing six-axis arm from ABB, Fanuc or Universal Robots, a 3D vision head typically mounted at the wrist, a suction or multi-finger gripper and a Covariant edge server. When an unseen SKU arrives in a tote, the Brain receives a point cloud plus RGB image from the camera, runs a single large transformer model trained on billions of pick attempts across the global Covariant fleet, outputs a grasp pose and a drop pose, and hands execution to the arm's own low-level motion planner. The model it runs since March 2024 is RFM-1, a multimodal foundation model trained on eight types of input (video of the operating cell, sensor readings, text instructions, URDF descriptions of the robot, tote metadata and more). Crucially, the Brain operates in a shared-learning mode: a grasp failure on an Obeta parcel-sortation cell in Berlin becomes a training example for an arm picking cosmetics in New Jersey the next morning. This "fleet learning" is the structural moat the company pitched for seven years.

Product and business overview

Covariant sold the Brain in three packaged configurations. Covariant Brain for Induction automates parcel induction into sorter conveyors and was the company's single biggest revenue line, deployed at Knapp, Pitney Bowes and ABB integrators. Covariant Brain for Each-Picking handled unit-level ecommerce order fulfilment out of totes, the use case Obeta ran on five arms. Covariant Brain for Depalletisation targeted mixed-SKU pallets, a harder problem the company only began to meaningfully commercialise in 2023. Alongside the three product lines it sold an annual AI-maintenance subscription that pushed model updates to deployed cells and retrained grasps for SKUs the fleet had not seen. Integration was done through named partners — ABB, Knapp, Fanuc and Fuji Yusoki — rather than Covariant's own field force, which kept gross margins palatable and head-count light but also handed the customer relationship to a hardware OEM.

Business model and pricing

Covariant never published list prices. Trade-press reporting and customer references place a single picking cell in the $200K–$400K range inclusive of arm, gripper, vision and the Brain, with Covariant capturing roughly $80K–$120K of that as software and hardware margin, plus an annual software subscription of $20K–$40K per cell for continuous model updates. By mid-2024 the company told investors it had around 100 live cells in production — a number that would imply an annual recurring revenue line in the low eight figures, consistent with the sub-$20M ARR figures that have leaked in trade press. The rest of revenue came from one-time integration fees and hardware pass-through.

Traction over time

Year	Milestone	Scale
2017	Founded as Embodied Intelligence	4 founders
2019	Rebrand to Covariant; first production pilot	<20 staff
2020	Obeta Berlin deployment goes live	~50 staff
2021	Series B at ~$625M valuation	~100 staff
2023	Series C at ~$750M; RFM-1 announced internally	~175 staff
2024	RFM-1 public launch (March); Amazon deal (August)	~180 staff → ~120 after deal

Reporters have estimated Covariant was running under $20M in ARR at the time of the Amazon deal, according to Bloomberg and GeekWire coverage dated 31 August 2024. That is a dramatically lower revenue base than the $400M-plus deal value would normally imply, and it is the single most revealing number on the page.

Market analysis

The warehouse-robotics market Covariant plays in is projected by Interact Analysis to hit roughly $63B by 2030, growing at a mid-teens compound rate as labour costs in US and European distribution centres compress operator margins. The specific wedge Covariant attacks — AI picking and sortation inside 3PLs, retailers and mail/parcel operators — is a narrower slice, around $8–10B in 2025. Structural tailwinds are real: US warehouse labour turnover ran above 45% annually through 2024 per the Bureau of Labor Statistics, parcel volumes are compounding at high single digits, and e-grocery each-pick is still largely done by humans. The headwind is equally real: hyperscalers (Amazon, Walmart, Meituan) are building their picking stacks in-house and increasingly open-sourcing pieces of them.

Competitive intel

Dexterity, based in Redwood City, raised a $95M Series C-1 in 2023 and is Covariant's closest technical peer on general-purpose picking; it attacks the same ABB/Fanuc integrators with similar AI claims and a slightly tighter focus on mixed-case palletising. Symbotic, now public at a market cap north of $25B, is a vertically integrated incumbent that sells an entire goods-to-person system to grocers like Walmart and Albertsons rather than a brain for someone else's arm; it is a direct threat for the retail accounts Covariant could not land. Berkshire Grey was a public-market darling until SoftBank took it private at a brutal discount in 2023 and folded its tech into its broader portfolio; its customer base is smaller than Covariant's but its hardware integration is tighter. Nimble Robotics and Fizyr compete on the software-only angle at a lower price point and with less compelling AI; Nimble skews to DTC brands, Fizyr to European logistics integrators. Amazon Robotics, now reinforced by the Covariant founding team, is the elephant: it does not sell to anyone but internalises every advance Covariant made.

History and evolution

Covariant's trajectory has three distinct phases. Phase one, 2017–2020, was pure research with a single flagship customer — a pattern reminiscent of early OpenAI. Phase two, 2020–2023, was the enterprise-sales-grind years: the company pushed hard into ABB integration, hired a VP of Sales from Blue Yonder in 2021 and attempted to turn fleet-learning into a reproducible deployment process. Growth was slower than investors wanted; a Series D reportedly aborted in mid-2023 when terms could not be agreed. Phase three, 2024, was the reverse acquihire: Amazon hired Abbeel, Chen and Duan and about 25% of the staff, took a non-exclusive licence to the foundation models, and left behind a 100-person services company. Ted Stinson, previously COO, stepped into the CEO role to run the remaining business. The pattern was the same one Microsoft used with Inflection and Google used with Character.AI: no FTC filing, no shareholder vote on an acquisition, but the founders, the IP and the talent all end up inside the acquirer.

What people say

The case for. Customers who ran Covariant in production — Obeta, Knapp integrators, Radial — consistently praised the Brain's ability to handle long-tail SKUs that competitors could not. Industry analysts at ARK Invest and Lux Research called the shared-fleet-learning architecture the single most impressive technical moat in warehouse robotics through 2023. Academic reviewers rated RFM-1's demonstration of language-conditioned manipulation as a genuine step change over prior work.

The complaints. Operators repeatedly flagged that integration cost ran 2–3x the sticker price of the cell, and that time-to-first-production often stretched past 12 months. Several European retail logistics partners said the Brain still required a bespoke model fine-tune for each new site, undercutting the fleet-learning story. Short-seller notes and LinkedIn reviews of ex-employees after the Amazon deal described a company that had burned through most of a $250M-plus capital stack without reaching $20M ARR and had been looking for a strategic buyer for 18 months.

Outlook: the open question

The decisive question is whether the residual Covariant can survive as a going concern selling foundation-model-powered picking to anyone who is not Amazon. For that to be true, the remaining engineering team must retain the ability to retrain RFM-1 without the three founders and their senior researchers (unlikely given the licence is non-exclusive but the know-how walked out the door), and Covariant must continue to win Knapp/ABB integrator deals against Dexterity on technical merit despite a brand that now reads "Amazon's cast-off". For it to be false, the shell liquidates into service revenue, customer accounts leak to Dexterity and Symbotic within 24 months, and Covariant becomes a case study rather than a company.

How to attack it

A new entrant with $15–30M in seed-plus-A capital can go directly at Covariant's remaining customer base with a software-only Brain that runs on any ABB, Fanuc or UR arm and charges per successful pick rather than per cell. The specific wedge is performance-based pricing: Covariant's $20K–$40K annual maintenance fee is a legacy of a world where customers paid for research hope; a modern entrant can price at $0.002–$0.005 per successful pick, aligned to the integrator's own throughput SLA, which both undercuts Covariant on cells with slow ramp and locks in the integrator's share-of-wallet. The biggest weaknesses the attacker can exploit are specific: integration cost (Knapp and ABB integrators have complained publicly to German trade press about month-long on-site tuning); retraining lag (adding a new SKU category typically takes 3–7 days in the field, per customer references in the German trade press); the thin bench after August 2024 (LinkedIn head-count data shows ~30% of senior research staff left in the three months following the Amazon deal); and brand damage (buyers now assume Covariant will be acquired again or wound down, pricing in migration risk). A credible attacker pairs an open-weights picking foundation model (building on OpenVLA or RT-X) with a hardware-partnership-first GTM, offloading field integration to the arm OEMs themselves. No ten-figure moat is needed; what is needed is a trust story that the IP and the team will still be there in three years.

Adjacent-segment play

The same core capability — a foundation model that converts raw sensor input into a grasp action — generalises most naturally to mixed-case depalletisation for food-service and beverage distribution, a market Dexterity already attacks and where PepsiCo, Sysco and US Foods are actively running pilots. It also generalises to kitting for medical-device and electronics contract manufacturers, a higher-margin segment where Fictiv (covered) sits adjacent and where cell throughput matters less than first-pass yield. A more interesting axis is DTC co-packing, where a $200K cell is too expensive but a $30K-per-month managed-service arm priced per shipment would be competitive against the 3PL labour cost; Nimble Robotics already attacks this and leaves room for a hardware-OEM-partnered entrant. The axis that does not work is consumer — humanoid picking in the home remains at least five years away from SKU-level reliability, and the Covariant team's own move into Amazon's internal fulfilment bets suggests they too concluded consumer was a decade too early.

Sources and further reading

• GeekWire, "Amazon hires Covariant founders, inks licensing deal with AI startup in latest 'reverse acquihire'", 31 August 2024, https://www.geekwire.com/2024/amazon-hires-covariant-founders-inks-licensing-deal-with-robotics-ai-startup-in-latest-reverse-acquihire-deal/
• TechCrunch, "Amazon hires the founders of AI robotics startup Covariant", 31 August 2024, https://techcrunch.com/2024/08/31/amazon-hires-the-founders-of-robotics-ai-startup-covariant/
• TechCrunch, "Covariant is building ChatGPT for robots", 11 March 2024, https://techcrunch.com/2024/03/11/covariant-is-building-chatGPT-for-robots
• BusinessWire, "Covariant Introduces RFM-1", 11 March 2024, https://www.businesswire.com/news/home/20240311948570/
• Supply Chain 24/7, "Amazon Hires Covariant Founders to Boost AI-Powered Robotics", 2 September 2024, https://www.supplychain247.com/article/amazon-hires-covariant-founders-boost-artificial-intelligence-powered-robotics
• Wikipedia, "Covariant (company)", accessed November 2026, https://en.wikipedia.org/wiki/Covariant_(company)
• Multiples.vc, "Covariant Acquisition, Valuation, Funding & Investors", https://multiples.vc/private-comps/covariant